In this paper we present an automated system that is able to track and grasp a moving object within the workspace of a manipulator using range images acquired with a Microsoft Kinect sensor. Realtime tracking is achieved by a geometric particle filter on the affine group. Based on the tracked output, the pose of a 7-DoF WAM robotic arm is continuously updated using dynamic motor primitives until a distance measure between the tracked object and the gripper mounted on the arm is below a threshold. Then, it closes its three fingers and grasps the object. The tracker works in real-time and is robust to noise and partial occlusions. Using only the depth data makes our tracker independent of texture which is one of the key design goals in our approach. An experimental evaluation is provided along with a comparison of the proposed tracker with state-of-the-art approaches, including the OpenNI-tracker. The developed system is integrated with ROS and made available as part of IRI's ROS stack.

In the present work, we propose and validate a complete probabilistic framework for human motion prediction in urban or social environments. Additionally, we formulate a powerful and useful tool: the human motion behavior estimator. Three different basic behaviors have been detected: Aware, Balanced and Unaware. Our approach is based on the Social Force Model (SFM) and the intentionality prediction BHMIP. The main contribution of the present work is to make use of the behavior estimator for formulating a reliable prediction framework of human trajectories under the influence of dynamic crowds, robots, and in general any moving obstacle. Accordingly, we have demonstrated the great performance of our long-term prediction algorithm, in real scenarios, comparing to other prediction methods.

We describe a system allowing a robot to learn goal-directed manipulation sequences such as steps of an assembly task. Learning is based on a free mix of exploration and instruction by an external teacher, and may be active in the sense that the system tests actions to maximize learning progress and asks the teacher if needed. The main component is a symbolic planning engine that operates on learned rules, defined by actions and their pre- and postconditions. Learned by model-based reinforcement learning, rules are immediately available for planning. Thus, there are no distinct learning and application phases. We show how dynamic plans, replanned after every action if necessary, can be used for automatic execution of manipulation sequences, for monitoring of observed manipulation sequences, or a mix of the two, all while extending and refining the rule base on the fly. Quantitative results indicate fast convergence using few training examples, and highly effective teacher intervention at early stages of learning.

In this talk I will present some recent advances on my research in active viewpoint planning for gathering data from an unexplored free-form and non-rigid 3D complex scenario. The aim of this research is to provide a general task-oriented approach based on information-gain maximization that easily deals with such a problem. In particular for this talk, I will show the advances on how to encode viewpoint planning in 3D occupancy models with the purpose of identifying available probing points in leaves. In overall terms, the approach consists of repetitively ranking a given set of possible viewpoints, based on their task-related gains, and then move to the best-ranked vantage point until the exploratory task termination criterion is achieved.

Not provided

In this talk we will first review which data provides an Inertial Measurement Unit (IMU), and how it is processed to be used in localization problems, focusing on aspects such as gravity cancellation and bias effects. Afterwards, we will present the work carried out in the context of the PipeGuard FP7 European project. In this later case IMU and encoder measurements were used to localize a sewer pipe inspection vehicle, following an iterative minimization over a sliding window of platform states. Experimental results will be shown, pointing out the strongest points of IMU devices, but analysing why they can't be alone in a localization application.

Advanced mobile robots today provide for the achievement of a variety of complex tasks, such as warehouse management (Kiva company), space exploration (NASA missions) or car manufacturing (industrial robots). The Spirit and Opportunity robots sent by NASA to Mars are such robots, evaluated at a staggering $625 million. Although an extreme case, with respect to cost, advanced mobile robots may be said to be, not only expensive, but also limited by a single point of failure i.e. one single robot.

The swarm robotic approach focuses on solving varied and complex tasks through collective behaviours emerging from the interaction between autonomous simple. Such agents are therefore simpler in their design and therefore more robust and affordable. Moreover the use of a large number of robots removes the single point of failure. However experiments in Swarm robotics are facing hardware and software challenges. On the hardware side, either the cost of a robot is prohibitive or the robot is limited in its functionalities. On the software side, it is difficult to design controllers for these robots, as a large number of interactions has to be taken into account.

I will present in more details the difficulties in designing a robot for swarm robotic and a solution developed at the CRAB Lab of NTNU. I will also present the difficulties in designing controller for these robots and the main avenues of research I'm exploring.

The combination of segmentation and recognition is a long-standing problem in computer vision. In this talk we will present an approach to exploit segmentation to construct appearance descriptors (e.g. SIFT) that can deal with occlusion and background changes, by downplaying measurements coming from pixels that are likely to belong to a different region than that of the descriptor's center. We will also present our current work, extending the same approach to object recognition with deformable part-based models, using SLIC superpixels to 'clean up' the HOG features and separate foreground and background.

Not disclosed.

A large number of robotic, computer vision and computer graphics applications rely on efficiently solving the associated sparse linear systems. Simultaneous localization and mapping (SLAM), structure from motion (SfM), non-rigid shape recovery, and elastodynamic simulations are only few examples in this direction. In general, these problems are nonlinear and the solution can be approximated by incrementally solving a series of linearized problems. In some applications, the size of the system considerably affects the performance, especially when the sparsity is low. Exploiting the block structure of such problems, we developed very efficient solutions to manipulate block matrices within iterative nonlinear solvers. We proposed a scheme which combines the advantages of block-wise schemes convenient in both, numeric and structural matrix modification and element-wise, which are efficient in arithmetic operations. SLAM++ is a very efficient implementation of incremental nonlinear least squares, based on the sparse block matrix scheme. It is aimed for use in 3D reconstruction and robotics. The implementation was designed with hardware acceleration in mind. This is very important for large scale problems, which can efficiently run on wide scale of accelerators, ranging from DSPs and FPGAs to clusters of GPUs.

Many situations in Robotics require an effective analysis of the motions of a closed-chain mechanism. Despite appearing very often in practice (e.g. in parallel manipulators, reconfigurable robots, or molecular compounds), there is a lack of general tools to effectively analyze the complex configuration spaces of such systems. In this talk I will give an overview of the Cuik Suite, an open-source toolbox for motion analysis of general closed-chain mechanisms. The package can determine the motion range of the whole mechanism or of some of its parts, detect singular configurations leading to control or dexterity issues, or find collision- and singularity-free paths between given configurations. The toolbox is the result of several years of research and development within the Kinematics and Robot Design group and is available under GPLv3 license.

Robots accompanying humans is one of the core capacities every service robot deployed in urban settings should have. We present a novel robot companion approach based on the so-called Social Force Model (SFM). A new model of robot-person interaction is obtained using the SFM which is suited for our robots Tibi and Dabo. Additionally, we propose an interactive scheme for robot's human-awareness navigation using the SFM and prediction information. Moreover, we present a new metric to evaluate the robot companion performance based on vital spaces and comfortableness criteria. Also, a multimodal human feedback is proposed to enhance the behavior of the system. The validation of the model is accomplished throughout an extensive set of simulations and real-life experiments.

This paper deals with the motion planning problem for parallel orienting platforms with one non-holonomic joint and two prismatic actuators which can maneuver to reach any three-degree-of-freedom pose of the moving platform. Since any system with two inputs and up to four generalized coordinates can always be transformed into chained form, this path planning problem can be solved using well-established procedures. Nevertheless, the use of these procedures requires a good understanding of Lie algebraic methods whose technicalities have proven a challenge to many practitioners who are not familiar with them. As an alternative, we show how by (a) properly locating the actuators, and (b) representing the platform orientation using Euler parameters, the studied path planning problem admits a closed-form solution whose derivation requires no other tools than ordinary linear algebra.

Graph edit distance is a powerful and flexible method for error-tolerant graph matching. Yet often it can only be calculated for small graphs due to its exponential time complexity. In this talk we propose a quadratic time approximation of graph edit distance based on the Hausdorff distance. In a series of experiments we analyze the performance of the proposed Hausdorff edit distance in the context of graph classification and compare it with a cubic time algorithm based on the assignment problem. Overall, the proposed Hausdorff edit distance shows a promising potential in terms of flexibility, efficiency, and accuracy.

The study of human-robot interaction (HRI) carries numerous challenges related to the concepts general to HRI and the specific use of robotic systems that interact with humans in a particular application domain. Mobile robots that operate in a home setting must be able to continuously track users and attend their calls for interaction. Recent release of affordable RGB-D sensors allowed development of new interaction modality that relies on real-time human body segmentation and can be applied in this domain. The proposed work, performed as a part of the EU-funded INTRO project, exploits 3D vision as input for intuitive interaction through person tracking, person following, gesture recognition and robot control. This seminar will introduce the algorithms and the challenges related to system integration and practical deployment of a RGB-D sensor on the autonomous mobile robot.

Most current depth sensors provide 2.5D range images in which depth values are assigned to a rectangular 2D array. In this paper we take advantage of this structured information to build an efficient shape descriptor which is about two orders of magnitude faster than competing approaches, while showing similar performance in several tasks involving deformable object recognition. Given a 2D patch surrounding a point and its associated depth values, we build the descriptor for that point, based on the cumulative distances between their normals and a discrete set of normal directions. This processing is made very efficient using integral images, even allowing to compute descriptors for every range image pixel in a few seconds. The discriminative power of our descriptor, dubbed FINDDD, is evaluated in three different scenarios: recognition of specific cloth wrinkles, instance recognition from geometry alone, and detection of reliable and informed grasping points.

Immense amounts of high resolution data are now routinely produced thanks to recent advances in EM imaging. While a strong demand for automated analysis now exists, it is stifled by the lack of robust automatic 3D segmentation techniques. State-of-the-art Computer Vision algorithms designed to operate on natural 2D images tend to perform poorly when applied to EM image stacks for a number of reasons. The sheer size of a typical EM image stack renders many segmentation schemes intractable. Most approaches rely on local statistics that easily become confused when confronted with the noise and textures found within EM image stacks. The assumption that strong image gradients always correspond to object boundaries is violated by cluttered membranes belonging to numerous objects. In this work, we propose an automated graph partitioning scheme that addresses these issues. It reduces the computational complexity by operating on supervoxels instead of voxels, incor- porates global shape features capable of describing the 3D shape of the target objects, and learns to recognize the distinctive ap- pearance of true boundaries. Our experiments demonstrate that, when applied to segment mitochondria from neural tissue, our approach closely matches the performance of human annotators and outperforms a state-of-the-art 3D segmentation technique.

Performing aerial 6-dimensional manipulation using flying robots is a challenging problem, to which only little work has been devoted. This paper proposes a motion planning approach for the reliable 6-dimensional quasi-static manipulation with an aerial towed-cable system. The novelty of this approach lies in the use of a cost-based motion-planning algorithm together with some results deriving from the static analysis of cable-driven manipulators. Based on the so-called wrench-feasibility constraints applied to the cable tensions, as well as thrust constraints applied to the flying robots, we formally characterize the set of feasible configurations of the system. Besides, the expression of these constraints leads to a criterion to evaluate the quality of a configuration. This allows us to define a cost function over the configuration space, which we exploit to compute good-quality paths using the T-RRT algorithm. As part of our approach, we also propose an aerial towed-cable system that we name the FlyCrane. It consists of a platform attached to three flying robots using six fixed-length cables. We validate the proposed approach on two simulated 6-D quasi-static manipulation problems involving such a system, and show the benefit of taking the cost function into account for such motion planning tasks.

This presentation is devoted to the real-time estimation of the pose (position and orientation in 3D space) of high-dynamic and naturally unstable robots such as aerial drones, humanoid robots, or ground robots running at very high speeds on rough terrain. The key point is the inclusion of inertial measurements fused with vision, leading to a quite elegant replica of what constitutes the oculo-vestibular system of superior animals, key for their sense and control of motion and equilibrium. The system draws on monocular, EKF-based simultaneous localization and mapping (SLAM). It was initially motivated by the need of real-time, rich and accurate pose estimates to be used within the control loop of the humanoid robot HRP2 at LAAS-CNRS, although it is equally useful in any device requiring fast pose updates with information of the gravity direction, crucial in naturally unstable robots of different nature. In particular, the outcome of the visual-inertial fusion algorithm is a vector of position, orientation, velocity, angular velocity, acceleration, gravity vector, and sensor biases, which is updated at frequencies ranging from 100Hz up to several kHz depending only on the IMU data-rate. These estimates are both globally and locally consistent thanks to the fusion of the complementary characteristics of the visual and inertial sensors used.

We introduce a novel approach to automatically recover 3D human pose from a single image. Most previous work follows a pipelined approach: initially, a set of 2D features such as edges, joints or silhouettes are detected in the image, and then these observations are used to infer the 3D pose. Solving these two problems separately may lead to erroneous 3D poses when the feature detector has performed poorly. In this paper, we address this issue by jointly solving both the 2D detection and the 3D inference problems. For this purpose, we propose a Bayesian framework that integrates a generative model based on latent variables and discriminative 2D part detectors based on HOGs, and perform inference using evolutionary algorithms. Real experimentation demonstrates competitive results, and the ability of our methodology to provide accurate 2D and 3D pose estimations even when the 2D detectors are inaccurate.

Hide-and-seek is a good game to study Human Robot Interaction; there are cognitive as well as engineering approaches to play the game as a robot. We have focused on the latter using different models and strategies to play as a seeker. Tests have been done extensively in simulations, and partly in real world experiments with Dabo. We introduce a new fast heuristic method for the seeker and a more complex method that uses MOMDP (Mixed Observability Markov Decision Processes) models to learn a good policy to be applied by the seeker. Even though MOMDPs reduce the computational cost of POMDPs (Partially Observable MDPs), they still have a high computational complexity which is exponential with the number of states. For the hide-and-seek game, the number of states is directly related to the number of grid cells. As an alternative to off-line MOMDP policy computation with the complete grid fine resolution, we have devised a two-level MOMDP, where the policy is computed on-line at the top level with a reduced number of states independent of the grid size.

Currently Robotics reached vehicular traffic through autonomous vehicles circulating several streets in the world. This advance was greatly influenced by the 2007 USA DARPA Grand Challenge where robot car teams participated from all world. Now, there are a new 2013-2014 Grand Challenge, but this time the leap will be on the Humanoid Robots. In this talk we will know about robotic trends and the principal players of the Robotics in the world through this big competition that could change the scope of the humanoids as we know.

This paper develops a new method for uncalibrated image-based visual servoing. In contrast to traditional image-based visual servo, the proposed solution does not require a known value of camera focal length for the computation of the image Jacobian. Instead, it is estimated at run time from the observation of the tracked target. The technique is shown to outperform classical visual servoing schemes in situations with noisy calibration parameters and for unexpected changes in the camera zoom. The method's performance is demonstrated both in simulation experiments and in a ROS implementation of a quadrotor servoing task. The developed solution is tightly integrated with ROS and is made available as part of the IRI ROS stack.

This paper presents a method to estimate external forces exerted on a manipulator during motion, avoiding the use of a sensor. The method is based on task-oriented dynamics model learning and a robust disturbance state observer. The combination of both leads to an efficient torque observer that can be incorporated to any control scheme. The use of a learning-based approach avoids the need of analytical models of joints' friction or Coriolis dynamics effects.

A [+]-manifold, just as a manifold, is a subspace S of R^s which can have a globally very complicated structure but locally behaves like R^n. An example is 3-D orientation. This structure is encapsulated in an operator [+]:S*R^n->S that allows to apply a local vector valued change to an element and an inverse operator [-]:S*S-->R^n. WIth these operators most estimation algorithms can work on [+]-manifolds, mostly by replacing + with [+] and - with [-].
The talk gives overview and intuition, why [+]-manifolds are a useful concept, how they are axiomatized and how they are used in estimation algorithms such as least-squares or UKF.

Recognizing manipulations performed by a human and the transfer and execution of this by a robot is a difficult problem. We address this by introducing a novel representation of the relations between objects at decisive time points during a manipulation. Thereby, we encode the essential changes in a visual scenery in a condensed way such that a robot can recognize and learn a manipulation without prior object knowledge. To achieve this we continuously track image segments in the video and construct a dynamic graph sequence. Topological transitions of those graphs occur whenever a spatial relation between some segments has changed in a discontinuous way and these moments are stored in a transition matrix called the (SEC). We demonstrate that these time points are highly descriptive for distinguishing different manipulations. Employing simple sub-string search algorithms, semantic event chains can be compared and type-similar manipulations can be recognized with high confidence. As the approach is generic, statistical learning can be used to find the archetypal SEC of a given manipulation class. The performance of the algorithm is demonstrated on a set of real videos showing hands manipulating various objects and performing different actions. In experiments with a robotic arm, we show that the SEC can be learned by observing human manipulations, transferred to a new scenario, and then reproduced by the machine.

Motion paths of cable-driven hexapods must carefully be planned to ensure that the lengths and tensions of all cables remain within acceptable limits, for a given wrench applied to the platform. The cables cannot go slack –to keep the control of the platform– nor excessively tight –to prevent cable breakage– even in the presence of bounded perturbations of the wrench. This talk presents a path planning method that accommodates such constraints simultaneously. Given two configurations of the platform, the method attempts to connect them through a path that, at any point, allows the cables to counteract any wrench lying inside a predefined uncertainty region. The resulting C-space is placed in correspondence with a smooth manifold, which allows defining a continuation strategy to search this space systematically from one configuration, until the second configuration is found, or path non-existence is proved by exhaustion of the search. The approach is illustrated on the NIST Robocrane hexapod, but it remains applicable to general cable-driven hexapods, either to navigate their full six-dimensional C-space, or any of its slices.

The discipline of pattern recognition is traditionally divided into the statistical and the structural approach. Statistical pattern recognition is characterized by representing objects by means of feature vectors, while the structural approach uses symbolic data structures, such as strings, trees, and graphs. In the current talk we focus on graphs for object representation. When comparing graph representations with feature vectors, one notices an increased flexibility and representational power provided by graphs. On the other hand, the domain of graphs lacks mathematical operations needed to build pattern recognition and machine learning algorithms and consequently, there is a shortage of suitable tools for graph classification, clustering, and related tasks.
In this talk, we review some advances in the field of graph-based pattern recognition that aim at making algorithmic tools originally developed in statistical pattern recognition available for graphs. We will focus on graph embedding and discuss relations with graph kernels. As a matter of fact, graph embedding provides us with an elegant and systematic way to make the complete arsenal of statistical pattern recognition tools available for graphs. The talk will also address computational complexity problems that arise with graph representations, discuss how they may possibly be overcome, and give an application example.

Obtaining object pose in low quality images cannot be achieved using the classical point features approach. To solve this cases we propose a method that learns on high quality controlled data the inner structure of the object. We will show that a strong understanding of the object makes possible recognition of it when bad or partial data is provided. To understand this inner structure we will use 3D data and high resolution images annotated with the pose of the object. Our main contributions are a new procedure to perform 3D training and a new boosting technique that is able to handle 3D visibility of the features. We compare against state-of-the-art methods for recognition in low quality data, we outperform them substantially.

In this talk I will present, at a high level, some recent advances on two fundamental graph matching problems: the graph edit distance and the computation of a median graph. In the first part, I will present a new method to compute the Graph Edit Distance between two attributed graphs which is based on Dominant Sets, a notion of cluster recently introduced. In the second part, I will show new theoretical results which link the problem of finding the Median Graph of a set of graphs to the Common Labelling problem. These are joint works with Albert Solé Ribalta, Francesc Serratosa and Marcello Pelillo.

Humanoids expected to collaborate with people should be able to interact with them in the most natural way. This involves significant perceptual, communication, and motor processes, operating in a coordinated fashion. Consider a social gathering scenario where a humanoid is expected to possess certain social skills. It should be able to explore a populated space, to localize people and to determine their status. Humans appear to solve these tasks routinely by integrating the often complementary information provided by multi sensory data processing.
In the talk will be presented some applications to achieve these goals. A sceneflow and a temporal version of a stereo seed growing algorithm are developed to provide to the robot information about the environment. In order to give to the robot some commands, an action recognition method based on the Bag-of-Words framework is implemented using a descriptor based on the 3D data and auditory information. Finally to make the interaction more real we would like that the robot identifies when a person is speaking and direct the head towards this person. This is achieved also using visual and auditory data fused by a Gaussian mixture model.

Current object class detectors learn models from generic images taken from different viewpoints and lighting conditions but using ground truth data. A lot of effort is usually devoted to manual labelling, since the goal is to generate a robust object model able to work for different appearances and aspects. Unfortunately, manual labelling is expensive, specially in long image sequences. Also, object detectors trained for general scenes usually exhibit poor performance under uncontrolled conditions. Finally, there are sequences where we are interested to detect particular objects for which there is not specific data, i.e. pre-trained detectors. In such cases, motion segmentation is still the most practical solution. In this paper, we propose an alternative solution based on a combination of motion segmentation and recent advances on weakly-supervised object detection. We propose a framework based on latent SVMs that, given a noisy initialization based on motion cues, is able to learn the correct position, the scale, and the appearance of all moving ob jects in the scene. We test our new approach on challenging sequences, and show that our detector outperforms motion segmentation as well as general object detectors.

The deployment of service robots in urban settings is an open issue for the robotics community. Tackling the problem by only relying on a robust navigation is not enough, although this topic has grown enormously in the few past years. Thus, the understanding of human motion in urban environments is of extreme importance in order to adapt service robots to typical human environments and not in the contrary. Human motion prediction in indoor and outdoor scenarios is a key issue towards an intelligent robot navigation and human robot interaction in general. In thiss seminar, I will present a new human motion intentionality indicator, denominated Bayesian Human Motion Intentionality Prediction (BHMIP), which is a geometric-based long-term predictor. The BHMP is compared with other long-time prediction state of the art methods using two well known databases. Additionally, experiments in a real scenario are carried out including a set of volunteers walking in the presence of a mobile two-wheeled robot, to validate the overall performance of the BHMIP.

We present an Online Random Ferns (ORFs) classifier that progressively learns and builds enhanced models of object appearances. During the learning process, we allow the human intervention to assist the classifier and discard false positive training samples. The amount of human intervention is minimized and integrated within the online learning, such that in a few seconds, complex object appearances can be learned. After the assisted learning stage, the classifier is able to detect the object under severe changing conditions. The system runs at a few frames per second, and has been validated for face and object detection tasks on a mobile robot platform. We show that with minimal human assistance we are able to build a detector robust to viewpoint changes, partial occlusions, varying lighting and cluttered backgrounds.

Active view planning for gathering data from an unexplored 3D complex scenario is a hard and still open problem in the computer vision community. In this paper, we present a general task-oriented approach based on an information-gain maximization that easily deals with such a problem. Our approach consists of ranking a given set of possible actions, based on their task-related gains, and then executing the best-ranked action to move the required sensor.
An example of how our approach behaves is demonstrated by applying it over 3D raw data for real-time volume modelling of complex-shaped objects. Our setting includes a calibrated 3D time-of-flight (ToF) camera mounted on a 7 degrees of freedom (DoF) robotic arm. Noise in the sensor data acquisition, which is too often ignored, is here explicitly taken into account by computing an uncertainty matrix for each point, and refining this matrix each time the point is seen again. Results show that, by always choosing the most informative view, a complete model of a 3D free-form object is acquired and also that our method achieves a good compromise between speed and precision.

We present an active exploration strategy that complements Pose SLAM and optimal navigation in Pose SLAM. The method evaluates the utility of exploratory and place revisiting sequences and chooses the one that minimizes overall map and path entropies. The technique considers trajectories of similar path length taking marginal pose uncertainties into account. An advantage of the proposed strategy with respect to competing approaches is that to evaluate information gain over the map, only a very coarse prior map estimate needs to be computed. Its coarseness is independent and does not jeopardize the Pose SLAM estimate. Moreover, a replanning scheme is devised to detect significant localization improvement during path execution. The approach is tested in simulations in a common publicly available dataset comparing favorably against frontier based exploration.

Motivated by the need of a robust and practical Inverse Kinematics (IK) algorithm for the WAM robot arm, we reviewed the most used closed-loop methods for redundant robots, analysing their main points of concern: convergence, numerical error, singularity handling, joint limit avoidance, and the capability of reaching secondary goals. As a result of the experimental comparison, we propose two enhancements. The first is to filter the singular values of the Jacobian matrix before calculating its pseudoinverse in order to obtain a more numerically robust result. The second is to combine a continuous task priority strategy with selective damping to generate smoother trajectories. Experimentation on the WAM robot arm shows that these two enhancements yield an IK algorithm that outperforms the reviewed state-of-the-art ones, in terms of the good compromise it achieves between time step length, Jacobian conditioning, multiple task performance, and computational time, thus constituting a very solid option in practice. This proposal is general and applicable to other redundant robots.

In certain configurations, usually referred to as singularities, the kinematic and static properties of mechanisms change dramatically. As these changes are often detrimental, the study of such singular configurations has been a major topic in the robotics literature. However, most of the work done concentrates on specific mechanisms and is motivated by the desire to avoid narrowly defined types of configurations. Even among specialists in robot kinetostatics, there is slight knowledge and often misunderstanding about what singularities are in general and what their effects can be. The talk will try to provide a simple but rigorous introduction to the topic. Particular attention will be given to lesser known types of singularity, exhibited by some parallel robots and other mechanisms with closed-loop kinematic chains.

In combinatorial characterization of rigidity one can simply count the vertices and edges (constraints) in a graph and its subgraphs to determine the rigidity and flexibility of a corresponding framework. The first complete combinatorial generalization for (generic) bar and joint frameworks in dimension 2 was confirmed by Laman in 1970. The 6|V|- 6 counting condition for 3-dimensional body-hinge and molecular structures, and a fast pebble game algorithm which tracks this count in the multigraph, have led to the development of the program FIRST, for rapid predictions of the rigidity/flexibility in proteins. In this talk we will show how our various extensions of the pebble game algorithm can be used to predict hinge motions in proteins, as well as some preliminary predictions for elusive allosteric communication - where binding on one portion of protein changes the shape or binding at a distance active site of the protein. We will demonstrate our rigidity-based allosteric communication on the important class of signalling proteins known as GPCRs. Time permitting, we will briefly discuss how our rigidity techniques and algorithms can be used to decompose a linkage into Assur graphs.

Neural circuits consist of neuronal arbors that synapse onto each other. The pattern of connectivity specifies the function that a circuit implements. Electron microscopy is currently the only imaging modality capable of unambiguously resolving the smallest circuit elements, namely individual synapses and terminal dendrites. Technological developments in electron microscopy, driven by the needs of neuroscience--large volumes at nanometer resolution--make possible the imaging of complete brains of small animals, and large chunks of mammalian nervous tissue. Even small animals like the Drosophila larva contain tens of thousands of neurons and millions of synapses, and therefore manual reconstruction is not possible. The stumbling block in extracting neural circuits from EM image volumes is the lack of fast and reliable algorithms for the reconstruction of neuronal arbors in 3d and the labeling of synaptic contacts. In this talk, I will show how neurons and synapses look like in EM image volumes; demonstrate all that goes wrong with current methods, and outline ideas for the development of algorithms that imitate humans in their ability to overcome ambiguities originating in noise and data loss for the successful reconstruction of neural circuits in desirable time scales.

This paper provides a method for computing force-feasible paths on the Stewart platform. Given two configurations of the platform, the method attempts to connect them through a path that, at any point, allows the platform to counteract any external wrench lying inside a predefined six-dimensional region. In particular, the Jacobian matrix of the manipulator will be full rank along such path, so that the path will not traverse the f orward singularity locus at any point. The path is computed by first characterizing the force-feasible C-space of the manipulator as the solution set of a system of equations, and then using a higher-dimensional continuation technique to explore this set systematically from one configuration, until the second configuration is found. Examples are included that demonstrate the performance of the method on illustrative situations.

This seminar will be a rehearsal of Nico's PhD defense.

We present a novel approach for the segmentation of 3-D point clouds into geometric surfaces using adaptive surface models. The intrinsic adaptation mechanism allows the segmentation to adapt to changing depth data, leading to stable, temporally coherent, and traceable segments. Starting from an initial configuration, the algorithm converges to a stable segmentation through a novel iterative split-and-merge procedure. By simply udpating the 3-D point cloud with the upcoming frame during the movie, the segmentation accommodates to the current data. This strategy further allows the formation of new segments during the video and the elimination of segments representing objects that are disappearing. We tested the method on a large variety of data acquired with different range imaging devices, including a structured-light sensor and a time-of-flight camera, and successfully segmented the videos into surfaces segments. We further demonstrate the feasibility of the approach using quantitative evaluations based on ground-truth data.

In this paper we present a novel method of designing multi-fingered robotic hands using tasks composed of both finite and infinitesimal motion. The method is based on representing the robotic hands as a kinematic chain with a tree topology. We represent finite motion using Clifford algebra and infinitesimal motion using Lie algebra to perform finite dimensional kinematic synthesis of the multi-fingered mechanism. This allows tasks to be defined not only by displacements, but also by the velocity and acceleration at different positions for the design of robotic hands. The additional information enables an increased local approximation of the task at critical positions, as well as contact and curvature specifications. An example task is provided using an experimental motion capture system and we present the design of a robotic hand for the task using a hybrid Genetic Algorithm/Levenberg-Marquadt solver.

In this talk we will introduce a class of passivity-based MRAC controllers that require Minimal Controller Synthesis (MCS). This means that, for systems in control canonical form with unknown (possibly time-varying) parameters and subject to matched, unmodelled nonlinearities and/or disturbances, the MCS strategy is able to ensure tracking of the states of a given reference model. We will also comment on the extensions of the MCS algorithm to discrete-time systems and its effectiveness in the implementation of MCS controllers for continuous-time plants. The proposal will be illustrated with numerical and experimental results as well.

We present a new approach to matching graphs embedded in R2 or R3. Unlike earlier methods, our approach does not rely on the similarity of local appearance features, does not require an initial alignment, can handle partial matches, and can cope with non-linear deformations and topological differences. To handle arbitrary non-linear deformations, we represent them as Gaussian Processes. In the absence of appearance information, we iteratively establish correspondences between graph nodes, update the structure accordingly, and use the current mapping estimate to find the most likely correspondences that will be used in the next iteration. This makes the computation tractable. We demonstrate the effectiveness of our approach first on synthetic cases and then on angiography data, retinal fundus images, and microscopy image stacks acquired at very different resolutions.

Many robotic hands use compliant joints because they provide several advantages when interacting with objects in unknown environments, but they also modify the relation between external and internal forces and vary the reachable workspace. We propose a detailed study of how compliant joints modify the statics of hands, from the point of view of parallel manipulators. Starting with fully actuated parallel manipulators, the chosen mathematical framework clarifies the role of joint compliance and allows to measure the reduction/increase of torque exerted by the active joints due to the influence of the passive compliant ones.

This paper provides an algorithm for computing singularity-free paths on non-redundant closed-chain manipulators. Given two non-singular configurations of the manipulator, the method attempts to connect them through a configuration space path that maintains a minimum clearance with respect to the singularity locus at all points. The method is resolution complete, in the sense that it always returns a path if one exists at a given resolution, or returns failure otherwise. The path is computed by defining a new manifold that maintains a one-to-one correspondence with the singularity-free configuration space of the manipulator, and then using a higher-dimensional continuation technique to explore this manifold systematically from one configuration, until the second configuration is found. Examples are included that demonstrate the performance of the method on illustrative situations.

Detecting grasping points is a key problem in cloth manipulation. Most current approaches follow a multiple re-grasp strategy for this purpose, in which clothes are sequentially grasped from different points until one of them yields to a desired configuration. In this paper, by contrast, we circumvent the need for multiple re-graspings by building a robust detector that identifies the grasping points, generally in one single step, even when clothes are highly wrinkled.
In order to handle the large variability a deformed cloth may have, we build a Bag of Features based detector that combines appearance and 3D geometry features. An image is scanned using a sliding window with a linear classifier, and the candidate windows are refined using a non-linear SVM and a grasp goodness criterion to select the best grasping point. We demonstrate our approach detecting collars in deformed polo shirts, using a Kinect camera. Experimental results show a good performance of the proposed method not only in identifying the same trained textile object part under severe deformations and occlusions, but also the corresponding part in other clothes, exhibiting a degree of generalization.

This work proposes a solution to the grasp synthesis problem, which consist of finding the best hand configuration to grasp a given object for a specific manipulation task while satisfying all the necessary constraints. This problem had been divided into sequential sub-problems, including contact region determination, hand inverse kinematics and force distribution, with the particular constraints of each step tackled independently. This may lead to unnecessary effort, such as when one of the problems has no solution given the output of the previous step as input. To overcome this issue, we present a kinestatic formulation of the grasp synthesis problem that introduces compliance both at the joints and the contacts. This provides a proper framework to synthesize a feasible and prehensile grasp by considering simultaneously the necessary grasping constraints, including contact reachability, object restraint, and force controllability. As a consequence, a solution of the proposed model results in a set of hand configurations that allows to execute the grasp using only a position controller. The approach is illustrated with experiments on a simple planar hand using two fingers and an anthropomorphic robotic hand using three fingers.

Face recognition has been a productive field of research for many years. Most of the existing solutions, however, are based on still images and only reach their best performance when the subjects want to be recognized. As face recognition systems continue to improve, better performance is demanded in extreme conditions or challenging situations, especially by the security industry, interested in more reliable biometric systems. Motion-based face recognition is a young research area which is motivated by this growing demand and is inspired by psychological studies that highlight the importance of motion when humans perform facial recognition. The talk will present a novel motion-based face recognition approach, the Expressive Deformation Profile, on which I've worked in collaboration with the Computer Vision Lab at the National University of Singapore. Results will be shown from experiments in a general Facial Expression Database (Cohn-Kanade) and in a twins' facial expression database to test the extreme challenge of distinguishing among identical-twins.

In this talk I will advocate the usage of the dichromatic reflection model for color image understanding. This model allows us to explain images by means of their intrinsic properties, and it allows us to discard scene incidental events such as specularities and shadows. The model separates the surface reflectance in Lambertian reflectance and specular reflectance.
I will apply the model to a variety of computer vision and image processing problems. The main aim of this talk is to show that when working with RGB values it is often beneficial to think in terms of the dichromatic reflection model. The model provides the link between variations of RGB values and physical events in the world. To demonstrate this, I will apply the model to derive photometric invariant differential structure of images. Next, I will use the model to estimate the illuminant in a scene, allowing for color constancy. The aim of the talk is to reveal the strength of the dichromatic model in color image understanding. Finally, I will conclude with several extensions of the model which include modeling multiple illuminants and ambient lighting.

Traditionally, the fields of learning and experimental design have been characterized for relying in convex functions, which allow to solve high dimensional problems efficiently. However, many scientific and engineering problems cannot be directly formulated in terms of convex functions. Classical approaches in learning and decision making are based on strong assumptions and approximations such as discretization of the search space or the selection of suboptimal (local) decisions. In this talk, I'm going to present some recent improvements in the fields of non-convex optimization, Bayesian optimization and non-parametric bandits which can be used as an efficient alternative to the convex/approximate solution. The talk includes a couple of robotics problems where these techniques have shown their potential. Finally, I'm going to present a free toolbox that we have recently developed with all these methods.

The problem of modeling paired input-output sequences arises in numerous application areas such as natural language processing, speech recognition and biology. One popular tool for modeling paired input-output sequences is to use Probabilistic Finite-State Transducers (FSTs). Because FSTs induce a latent state representation they can be very effective in modeling the intrinsic dynamics of input-output sequences. My recent work has focused on studying how FSTs can be used to solve real sequence prediction tasks. This involves developing efficient learning algorithms for inducing FSTs from supervised data.
Most algorithms for FST learning rely on gradient-based or EM optimizations which can be computationally expensive and suffer from local optima issues. In the first part of the talk, I will follow a different approach and present a spectral algorithm for FST learning with strong PAC-style guarantees. At its core, the algorithm is simple and scalable to large data sets. Our approach follows a both recent and no-so-recent line of work which is based on using multiplicity automaton representations of finite state machines. In the second part of the talk, I will discuss ongoing work on using this algorithm for sequence prediction problems in natural language processing and highlight possible uses of FST learning in other application areas.

The inverse kinematics (IK) of a robot is the mapping that, given a goal position, calculates a set of joint positions so as to place the robot's end effector in the specified goal. As this is one of the most relevant computations to move the robot, there has been a lot of work about obtaining a fast and robust IK algorithm. In this seminar we will explain the difficulties on analytically obtaining the inverse kinematics of a robot, and we present the most used closed-loop methods, pointing out their main concerns, numerical error, convergence, capability of reaching secondary goals, and respecting joint limits. We also present two enhancements of some of the methods exposed. One of them, by filtering the Jacobian matrix eigenvalues before calculating its pseudoinverse. The other, as a combination of other existing methods.

A learning framework with a bidirectional communication channel is proposed, where a human performs several demonstrations of a task using a haptic device (providing him/her with force-torque feedback) while a robot captures these executions using only its force-based perceptive system. Our work departs from the usual approaches to learning by demonstration in that the robot has to execute the task blindly, relying only on force-torque perceptions, and, more essential, we address goal-driven manipulation tasks with multiple solution trajectories, whereas most works tackle tasks that can be learned by just finding a generalization at the trajectory level. To cope with these multiple-solution tasks, in our framework demonstrations are represented by means of a Hidden Markov Model (HMM) and the robot reproduction of the task is performed using a modified version of Gaussian Mixture Regression that incorporates temporal information (GMRa) through the forward variable of the HMM. Also, we exploit the haptic device as a teaching and communication tool in a human-robot interaction context, as an alternative to kinesthetic-based teaching systems. Results show that the robot is able to learn a container-emptying task relying only on force-based perceptions and to achieve the goal from several non-trained initial conditions.

Not provided by the speaker.

Whenever service robots have to interact with humans they need to perceive the environment in a similar way people do. This talk presents several works related to robotic perception in indoor environments. It first introduces the categorization of typical places found indoors, like for example corridors and offices, using different types of sensors. In addition, it presents some techniques to detect other agents and objects such as people or typical pieces of furniture. Finally, it describes a conceptual representation of an indoor environment that can be created by a service robot by integrating the previous approaches.

Existing sampling-based robot motion planning methods are often inefficient at finding trajectories for kino- dynamic systems, especially in the presence of narrow passages between obstacles and uncertainty in control and sensing. To address this, we propose EG-RRT, an Environment-Guided variant of RRT designed for kinodynamic robot systems that combines elements from several prior approaches and may incorporate a cost model based on the LQG-MP framework to estimate the probability of collision under uncertainty in control and sensing. We compare the performance of EG-RRT with several prior approaches on challenging sample problems. Results suggest that EG-RRT offers significant improvements in performance.

In many relevant path planning problems, loop closure constraints reduce the configuration space to a manifold embedded in the higher-dimensional joint ambient space. Whereas many progresses have been done to solve path planning problems in the presence of obstacles, only few work consider loop closure constraints. In this paper, we present the AtlasRRT algorithm, a planner specially tailored for such constrained systems that builds on recently developed tools for higher-dimensional continuation. These tools provide procedures to define charts that locally parametrize manifolds and to coordinate them forming an atlas. AtlasRRT simultaneously builds an atlas and a Rapidly-Exploring Random Tree (RRT), using the atlas to sample relevant configurations for the RRT, and the RRT to devise directions of expansion for the atlas. The advantage of the new planner comes from that samples obtained from the atlas allow a more efficient extension of the RRT as compared to state of the art approaches where samples are generated in the joint ambient space.

In general, high-order coupler curves of plane mechanisms cannot be properly traced by standard predictor-corrector algorithms due to drifting problems and the presence of singularities. Instead of focusing on finding better algorithms for tracing curves, a simple coordinate-free method that first traces these curves in a distance space and then maps them onto the mechanism workspace is proposed. Tracing a coupler curve in the proposed distance space is much simpler because (a) the equation of this curve in this space can be straightforwardly obtained from a sequence of bilaterations; and (b) the curve in this space naturally decomposes into branches in which the signs of the oriented areas of the triangles involved in the aforementioned bilaterations remain constant. A surjective mapping permits to map the thus traced curves onto the workspace of the mechanism. The advantages of this two-step method are exemplified by tracing the coupler curves of a double butterfly linkage, curves that can reach order 48.

Dr. Reis Simmons is a professor at de Carnigie Mellon University, his research interests focus on developing reliable, highly autonomous systems (especially mobile robots) that operate in rich, uncertain environments. The goal is to create intelligent systems that can operate autonomously for long periods of time in unstructured, natural environments. This necessitates robots that can plan, effectively reason about uncertainty, diagnose and recover from unanticipated errors, and reason about their limitations. In particular, he is interested in architectures for autonomy that combine deliberative and reactive behavior, reliable execution monitoring and error recovery, multi-robot coordination, probabilistic and symbolic planning, formal verification of autonomous systems, and human-robot social interaction.

This seminar will present how to make improved action selection for online policy learning in robotic scenarios using reinforcement learning (RL) algorithms. Since finding control policies using any RL algorithm can be very time consuming, we combine RL algorithms with heuristic functions for selecting promising actions during the learning process. Several methods have been successfully applied for defining the heuristic function, including defining a heuristics using information from the learning process, the use of previous domain knowledge in fixed heuristics and the reuse of previously learned policies, in a Case-Based Reasoning approach. Experimental results on robot navigation will show that the use of very simple heuristic functions results in significant performance enhancement of the learning rate.

Since it was proposed in 1989, occupancy grids have been a very popular technique to represent the environment. Despite its efficiency, occupancy grids have a number of drawbacks - it requires an arbitrary discretisation of the space, it makes strong assumptions of independence between cells, and it does not scale well to 3D, typically requiring external data structures. In this talk I will introduce a novel methodology to continuous occupancy maps directly addressing these issues, the Gaussian Process Occupancy Maps (GPOM). The problem of occupancy mapping is tackled as a classification task where the robot's environment is classified into regions of occupancy and free space. This is obtained by employing a modified Gaussian process as a non-parametric Bayesian learning technique to exploit the fact that real-world environments inherently possess structure. The result is an anytime algorithm capable of generating accurate representations of large environments at arbitrary resolutions to suit many applications. I will show how to extend GPOM to account for the vehicle's uncertainty in position, and efficient strategies to handle 3D representations.

Nowadays, research opens new fields in the development of entertainment robots, those robots are able to interact with people and share their recreation time. Following this idea, Dr. Reid Simmons, Professor of the Robotics Institute at the Carnigie Mellon University (CMU), began a project in which a robot plays the scrabble game with 1, 2 or 3 people. Besides, the robot is able to interact with people through dialogue and facial expressions. In consequence, this seminar will be focused on the dialogue of the robot to create an engagement with people and start the game, furthermore, it will be shown how should be the conversation during the game depending on players' behavior.

Se desarrollarán temas de fondo en el área de la Identificación y Control de sistemas dinámicos. Se expondrán los fundamentos de la Identificación Robusta y de la Falsificación de Controladores surgidas a principios y finales de los '90, respectivamente. Se tratará especialmente el manejo de la información relativa a la determinación de familias de modelos, la invalidación de los mismos, los niveles de conservadurismo (u optimismo) y su uso en el control de sistemas físicos. Se ejemplificará con casos de aplicación en los que intervino el autor.

This talk presents a procedure to optimize the quality of robotic grasps for objects that need to be held and manipulated in a specific way, characterized by a number of tight contact constraints. The main difficulties of the problem include that the set of feasible grasps is a manifold implicitly defined by a system of non-linear equations, the high dimension of this manifold, and the multi-modal nature of typical grasp quality indices, which make local optimization methods get trapped into local extrema. The proposed procedure finds a way around these difficulties by focussing the exploration on a relevant subset of grasps of lower dimension, which is traced out exhaustively using higher-dimensional continuation techniques. Using these techniques, a detailed atlas of the subset is obtained, on which the highest quality grasp according to any desired criterion can be readily identified. We will show results on a 3-finger planar hand and on the Schunk anthropomorphic hand to validate the approach.

We present an experimental evaluation of Boosted Random Ferns in terms of the detection performance and the training data. We show that adding an iterative bootstrapping phase during the learning of the object classifier, it increases its detection rates given that additional positive and negative samples are collected (bootstrapped) for retraining the boosted classifier. After each bootstrapping iteration, the learning algorithm is concentrated on computing more discriminative and robust features (Random Ferns), since the bootstrapped samples extend the training data with more difficult images. The resulting classifier has been validated in two different object datasets, yielding successful detections rates in spite of challenging image conditions such as lighting changes, mild occlusions and cluttered background.

In this talk we present some recent theoretical developments concerning the following topics related with energy applications:

Simultaneous Localization and Mapping (SLAM) and Structure from Motion (SFM) are important and closely related problems in robotics and vision. Not surprisingly, there is a large literature describing solutions to each problem, and more and more connections are established between the two fields. At the same time, robotics and vision researchers alike are becoming increasingly familiar with the power of graphical models as a language in which to represent inference problems. In this talk I will show how SLAM and SFM can be posed in terms of factor graphs, and how inference in them can be explained in a purely graphical manner via the concept of variable elimination. I will also show that, when applied to Gaussian problems, the algorithm yields the familiar QR and Cholesky factorization algorithms, and that this connection with linear algebra leads to strategies for very fast inference in arbitrary graphs. I will then present the Bayes tree as a novel data structure for representing the inferred posteriors. The Bayes tree is similar to a junction tree, yet better embodies the connection with sparse linear algebra. I will conclude by showing some published and preliminary work that exploits this connection to the fullest.

In last years many research work has focused in different areas within multi-robot systems. One of these areas is multi-robot task allocation, which involves the distribution of tasks between robots. Despite the large number of proposals in literature, there are already many open issues inside multi-robot task allocation, for example task reallocation.
Task reallocation or dynamic task allocation, implies the capacity of the system to change previously done assignations (links between robots and tasks) to get "best" system performance.
In this paper we present a distributed task reallocation approach, that relies on an energy-based formalization which considers tasks that change over time. We use simulation to show the performance of this proposal and to compare with other architectures.

The project developed uses a single camera to capture images of a moving human hand. The position and orientation of the finger nails is estimated throughout time, using computer vision. A dimensional kinematic synthesis solver is used in order to calculate the joint axes and the rotations of around each joint for a task. This allows creating personalized kinematic models of human hands for use in both prosthesis and exoskeleton designs.

Supervision of long-lasting extensive botanic experiments is a promising robotic application that some recent technological advances have made feasible. Plant modelling for this application has strong demands, particularly in what concerns 3D information gathering and speed. This paper shows that Time-of-Flight (ToF) cameras achieve a good compromise between both demands, providing a suitable complement to color vision. A new method is proposed to segment plant images into their composite surface patches by combining hierarchical color segmentation with quadratic surface fitting using ToF depth data. Experimentation shows that the interpolated depth maps derived from the obtained surfaces fit well the original scenes. Moreover, candidate leaves to be approached by a measuring instrument are ranked, and then robot-mounted cameras move closer to them to validate their suitability to being sampled. Some ambiguities arising from leaves overlap or occlusions are cleared up in this way. The work is a proof-of-concept that dense color data combined with sparse depth as provided by a ToF camera yields a good enough 3D approximation for automated plant measuring at the high throughput imposed by the application.

The probabilistic belief networks that result from standard feature-based simultaneous localization and map building cannot be directly used to plan trajectories. The reason is that they produce a sparse graph of landmark estimates and their probabilistic relations, which is of little value to find collision free paths for navigation. In contrast, we argue in this paper that Pose SLAM graphs can be directly used as belief roadmaps. We present a method that devises optimal navigation strategies by searching for the path in the pose graph with lowest accumulated robot pose uncertainty, independently of the map reference frame. The method shows improved navigation results when compared to shortest paths both over synthetic data and real datasets.

In this talk, several control problems for fuel cell systems will be studied; specifically I will focus on the temperature and air supply control for stand-alone fuel cell systems and on the power management for fuel cell hybrid vehicles. For both cases, advanced MPC formulations (hybrid and explicit formulations) will be proposed to deal with piece-wise affine models, binary variables and short sampling times. The advantages of this technology and real-time implementation feasibility will be discussed with experimental data.

Nowadays, the solution of the position analysis problem of a planar linkage almost invariably implies, as a first step, obtaining a system of equations from the independent kinematic loops of the mechanism, since general methods for the analysis of any planar linkage has been developed based on this set of equations in the last 10-15 years. As a consequence, the use of this approach, the so-called loop closure equations method, has been so overwhelming that other approaches for deriving constraints of the valid configurations of a linkage seem to have been banished from literature. However, this standard approach is beginning to be questioned because even for simple linkages the resulting system of kinematic equations is so involved that its solution requires the use of complex variable elimination techniques or special numerical algorithms. In this seminar, the modular kinematic analysis of planar linkages using Baranov trusses is presented as an alternative to the loop closure equations method. Basic definitions, advantages, drawbacks and examples of this approach are discussed.

The history of adaptive control systems dates back to early 50-s, when the aeronautical community was struggling to advance aircraft speeds to higher Mach numbers. In November of 1967, X-15 launched on what was planned to be a routine research flight to evaluate a boost guidance system, but it went into a spin and eventually broke up at 65,000 feet, killing the pilot Michael Adams. It was later found that the onboard adaptive control system was to be blamed for this incident. Exactly thirty years later, fueled by advances in the theory of nonlinear control, Air Force successfully flight tested the unmanned unstable tailless X-36 aircraft with an onboard adaptive flight control system. This was a landmark achievement that dispelled some of the misgivings that had arisen from the X-15 crash in 1967. Since then, numerous flight tests of Joint Direct Attack Munitions (JDAM) weapon retrofitted with adaptive element have met with great success and have proven the benefits of the adaptation in the presence of component failures and aerodynamic uncertainties. However, the major challenge related to stability/robustness assessment of adaptive systems is still being resolved based on testing the closed-loop system for all possible variations of uncertainties in Monte Carlo simulations, the cost of which increases with the growing complexity of the systems. This talk will give an overview of the limitations inherent to the conventional adaptive controllers and will introduce the audience to the L1 adaptive control theory, the architectures of which have guaranteed robustness in the presence of fast adaptation. Various applications, including flight tests of a subscale commercial jet, will be discussed during the presentation to demonstrate the tools and the concepts. With its key feature of decoupling adaptation from robustness L1 adaptive control theory has facilitated new developments in the areas of event-driven adaptation and networked control systems. A brief overview of initial results and potential directions will conclude the presentation.

This talk is mainly divided in two parts. The first one presents a study of motion segmentation problems. Based on this study, a novel algorithm for mobile-object segmentation from a static background scene is also presented. This approach is demonstrated robust and accurate under most of the common problems in motion segmentation. The second one tackles the problem of shadows in depth. Firstly, a bottom-up approach based on a chromatic shadow detector is presented to deal with umbra shadows. Secondly, a top-down approach based on a tracking system has been developed in order to enhance the chromatic shadow detection.
In our first contribution, a case analysis of motion segmentation problems is presented by taking into account the problems associated with different cues, namely colour, edge and intensity. Our second contribution is a hybrid architecture which handles the main problems in the case analysis, by fusing (i) these three cues and (ii) a temporal difference algorithm. On the one hand, we enhance the colour and edge models to solve both global/local illumination changes (shadows and highlights) and camouflage in intensity. In addition, local information is exploited to cope with camouflage in chroma. On the other hand, the intensity cue is also applied when colour and edge cues are not available, such as when beyond the dynamic range. Additionally, temporal difference is included to segment motion when these three cues are not available. Lastly, the approach is enhanced for allowing ghost detection.
Most segmentation approaches dealing with shadow detection are typically restricted to penumbra shadows. Therefore, such techniques cannot cope well with umbra shadows.
In the Bottom-up part, the shadow detection approach applies a novel technique based on gradient and colour models for separating chromatic moving shadows from moving objects. Hereafter, the regions corresponding to potential shadows are grouped by considering a bluish effect and an edge partitioning. Lastly, (i) temporal similarities between local gradient structures and (ii) spatial similarities between chrominance angle and brightness distortions are analysed for all potential shadow regions in order to finally identify umbra shadows. In the top-down process, after detection of objects and shadows both are tracked using Kalman filters. Firstly, this implies a data association and a case analysis between the blobs (foreground and shadow) and Kalman filters. Based on this association, temporal consistency is looked for the association between foregrounds and shadows and their respective Kalman Filters. From this association several cases are studied, as a result lost chromatic shadows are correctly detected. As a result, our approach obtains very accurate and robust motion segmentation in both indoor and outdoor scenarios, as quantitatively and qualitatively demonstrated in the experimental results, by comparing our approach with most best-known state-of-the-art approaches.

Robot learning from demonstration stands out as a promising way to endow robots with learning capabilities and allow human beings to teach them tasks in a natural way. This talk addresses how a robot can learn a multi-trajectory task based on its force/torque perceptions while a human teacher demonstrates the task, and how he/she can be more involved in the learning process by feeling through haptic feedback what the robot senses during the demonstration phase. Here, a learning framework with a bidirectional communication channel is proposed, where a human teacher teleoperates a robotic arm using a haptic interface. The teacher demonstrates a container-emptying task while receiving F/T feedback sensed by a force sensor placed on the robotic wrist. Teacher demonstrations are analyzed under the Mutual Information criterion for extracting relevant input variables for the task. Resulting data streams are statistically encoded using a Hidden Markov Model (HMM) and the execution phase is developed by implementing a modified version of Gaussian Mixture Regression that uses implicit temporal information from the HMM. Results show how the robot is able to learn a multi-trajectory task using only force-based perceptions, which can be exploited in further settings where other perception systems do not provide enough information about a specific task to be learned or scenarios where force-based data provide a better understanding about what the robot has to learn.

Attributes have recently shown to give excellent results for category recognition. In this paper, we demonstrate their performance in the context of image retrieval. First, we show that retrieval based on attribute vectors gives results comparable to the state of the art if retrieving images of a particular object. Second, we demonstrate that combining attribute and Fisher vectors improves performance for retrieval of the same object as well as categories. Third, we implement an efficient coding technique for compressing the combined descriptor to very small codes. Experimental results on the Holidays dataset show that our approach significantly outperforms the state-of-the-art, even for a very compact representation of 16 bytes per image. Retrieving category images is evaluated on a web dataset set with text annotation for learning representations. We present a baseline for retrieval based on images only and show the contribution of attribute features. Furthermore, we show how the combined image features can supplement text features.

The functioning and development of living organisms is controlled on the molecular level by networks of genes, proteins, small molecules, and by their mutual interactions. Prediction, control, and understanding of these systems, named gene regulatory networks, arise mainly from modelling them using iterated computer simulations and non-linear mathematical analysis. Biotechnological advances in quantitative high-throughput technology, in combination with the growing inter-disciplinarity between biology with engineering and natural sciences, have made this challenge achievable thanks to the emerging fields of Systems and Synthetic biology. In the seminar we will first present an introduction to the mentioned disciplines, and to the mathematical tools commonly used. Secondly, we will provide, as example, results about the mathematical modelling and non-linear analysis of IRMA, a novel synthetic network built in the yeast for benchmarking different computational methods.

The types of tasks that a manipulator can complete have been taken one step further using machine learning (ML) approaches. Tasks where the robot interacts with very complex dynamics have been used to demonstrate the benefits of ML in the exploitation of manipulator capacities. This seminar will review some state-of-the-art techniques in ML for the generation of robotic motions, including approaches based on learning by demonstration and reinforcement learning in continuous states and action spaces.

It will be described the modeling and model predictive control of sewage water networks in presence of episodes of heavy rain. The results are based on a new branching algorithm that has been developed and implemented in the last months in Heidelberg University to solve the associated optimization problem.

Generalized Voronoi Diagram (GVD) is an important geometrical technique with several applications in different fields. In this talk it will be discussed an application of GVD to autonomous robot path planning with a global vision system. The GVD technique was implemented and some experiments were conducted with Khepera Lab Robots.

El present projecte és un estudi sobre un aspecte important en el disseny d'un robot paral.lel. En particular, l'objectiu és desenvolupar i aplicar una metodologia per a trobar les posicions d'uns punts característics d'una plataforma de Gough-Stewart reconfigurable. Específicament, els punts a trobar del robot són els punts d'ancoratge de les cadenes cinemátiques tant a la base com a la plataforma, que fa d'efector final. Un objectiu secundari és determinar si és factible l'ús de servo motors de mida i cost reduïts per a la fabricació de robots paral.lels. Per aconseguir aixó s'ha desenvolupat un métode numéric i s'ha construït un prototipus sobre el que s'ha provat aquest métode. Degut a la complexitat de definir matemàticament la naturalesa dels robots paral.lels i al seu elevat nombre d'incògnites, el cálcul utilitzat és un algorisme iteratiu basat en el métode de Newton. Els resultats d'aplicar aquesta metodologia són prou positius. No obstant, Les posicions trobades són tan properes com ho permet la precisió dels seus elements, ja sigui la del sensor com la dels propis motors.

Se presentará un nuevo método no-iterativo para la obtención de la pose de la cámara y la distancia focal a partir de correspondencias 2D-3D. El método presentado ha demostrado ser robusto a ruido y a correspondencias erróneas. Se ha utilizado una aproximación similar a la utilizada en el algoritmo EPnP. Mostramos que las técnicas de linearización y relinearización utlizadas en el algoritmo EPnP presentan limitaciones que hemos solventado mediante lo que hemos definido como linearización y relinearización exhaustiva.
Presentaremos resultados sobre datos reales y sintáticos; a la vez, compararemos con métodos totalmente calibrados, el caso del EPnP, y con métodos como el DLT que necesitan encontrar la distancia focal. Estos resultados corroboran que el método presentado es capaz de ofrecer una precisión y robustez superior a otros métodos no calibrados y comparable a métodos calibrados.

Despite the significant advances in path planning methods, problems in highly constrained spaces are still challenging. In particular, in many situations the configuration space is a non-parametrizable variety implicitly defined by constraints, which complicates the successful generalization of sampling-based path planners. In this paper, we present a new path planning algorithm specially tailored for highly constrained systems. It builds on recently developed tools for Higher-dimensional Continuation, which provide numerical procedures to describe an implicitly defined variety using a set of local charts. We propose to extend these methods to obtain an efficient path planner on varieties, handling highly constrained problems. The advantage of this planner comes from that it directly operates into the configuration space and not into the higher-dimensional ambient space, as most of the existing methods do.

En este trabajo se presentan herramientas interactivas en 2D y 3D para ilustrar los conceptos teóricos de control en modo de deslizamiento de sistemas dinámicos lineales a trozos. El objetivo final es permitir a los alumnos la posibilidad de visualizar la evolución del sistema y su respuesta en tiempo real, ante los cambios y configuraciones de toda una serie de parámetros del sistema en estudio. De esta manera, podrán obtener conocimientos prácticos y avanzar en la comprensión de la respuesta del sistema de manera cualitativa y cuantitativa. Las herramientas se han desarrollado mediante Sysqueke y Easy Java Simulations (EJS), que permite una fácil programación y diseño.

In this seminar we are going to discuss the existence conditions of cuspidal configurations on some parameters of a special degenerate 3-RPR parallel manipulator. These configurations will be detected as cusp points, which make possible non-singular changes of assembly-mode and thus increase the singularity-free workspace.
We will combine algebraic-geometric techniques, such as Gröbner bases, discriminant varieties and cylindrical algebraic decomposition, in order to obtain a distribution of the number of cusp points. This distribution will let us analyze a whole family of degenerate 3-RPR manipulators that depend on one (or some) geometric parameter.
As a direct application of the presented results we obtain a simple method to determine the values of either the geometric parameters or the leg-rod lengths that make the manipulator be non-cuspidal.

The correspondence problem in computer vision is of basic importance since it arises at the early stages of many computer vision applications. So, it is of basic importance to develop efficient methods that are both robust -in the sense of being able to deal with noisy measurements- and, general -in the sense of having a wide field of application-. Attempts to solve the correspondence problem are based either on correlation or on feature descriptors. Since these methods use the local information around the interest points, that may produce erroneous matches under certain circumstances. Outlier rejectors such as RANSAC or Graph Transformation are used remove these erroneous matches by using global geometrical information. Since these outlier rejectors are unable to modify or produce new correspondences (different from those in the initial tentative-set), this may lead to poor results in the case of bad initial tentative-sets. Attributed graph matching methods are optimization techniques that exploit the nodes' attributes and their structural relations to compute new correspondences. We present an attributed graph matching approach that uses the geometrical information of the feature points in an affine-invariant way. It is also able to detect outlying features in both sides of the assignment, unlike other graph matching methods in the literature. We evaluate its performance in image matching and shape retrieval experiments.

This seminar introduces methods for Large scale SLAM developed during my stay at Georgia Tech. The motivation of this work comes from the fact that SLAM is a large optimization problem which requires very fast solvers when the result is used in real robotic applications. I will introduce two concepts; the Bayes tree and the subgraph preconditioning conjugate gradients (SPCG) and will conclude with an application to underwater 3D reconstruction of submerged structures.
The Bayes tree is a novel data structure that provides an algorithmic foundation enabling a better understanding of existing graphical model inference algorithms and their connection to sparse matrix factorization methods. Similar to a clique tree, a Bayes tree encodes a factored probability density, but unlike the clique tree it is directed and maps more naturally to the square root information matrix of the SLAM problem. The Bayes tree is used to obtain a completely novel algorithm for sparse nonlinear incremental optimization, that combines incremental updates with fluid relinearization of a reduced set of variables for efficiency. The new algorithm insures fast convergence to the exact solution.
SPCG is obtained by re-interpreting the method of conjugate gradients in terms of the graphical model representation of the SLAM problem. The main idea is to combine the advantages of direct and iterative methods, by identifying a sub-problem that can be easily solved using direct methods, and solving for the remaining part using PCG. The easy sub-problems correspond to a spanning tree, a planar subgraph, or any other substructure that can be efficiently solved. As such, our approach provides new insights into the performance of state of the art iterative SLAM methods based on re-parametrized stochastic gradient descent.

The dynamic response of PEM fuel cells has a relatively slow component, mainly due to water and reactant gases dynamics. To overcome this limitation, FC-supercapacitors (SCs) topologies can be used in order to manage very fast power variations, presenting in addition high power density, long lifecycle and good charge/discharge efficiency. In this seminar, a FC\SC-based autonomous hybrid system for residential applications that is currently under development in the ACESs Fuel Cell Laboratory is presented. The FC and SCs are connected in parallel, through two separate DC/DC converters, to a DC bus. Under steady state conditions, the FC must deliver the load power requirement, while maintaining the SCs voltage regulated to the desired value. Under sudden load variations, the FC current rate must be limited to assure a safe transition to the new point of operation. During this current rate limitation mode, the SCs must deliver or absorb the power difference. To this end, a sliding mode strategy is proposed to satisfy to control objectives. The main one is the robust regulation of the DC bus voltage, even in the presence of system uncertainties and disturbances, such as load changes and FC voltage variations. Additionally, a second control objective is attained, namely to guarantee the adequate level of charge in the SCs, once the FC reaches the new steady state operation point. In this way, the system can meet the load power demand, even under sudden changes, and it can also satisfy a power demand higher than the nominal FC power, during short periods. The preliminary results presented, corresponds to the first stage of a joint project for the design and development of a novel FC\SCs-based autonomous hybrid system.

This talk provides a detailed description of a set of algorithms to efficiently manipulate 3D geometric models to compute physical constraints and range observation models, data that is usually required in real-time mobile robotics or simulation. Our approach uses a standard file format to describe the environment and processes the model using the openGL library, a widely-used programming interface for 3D scene manipulation. The talk also presents results on a test model for benchmarking, and on a model of a real urban environment, where the algorithms have been effectively used for real-time localization in a large urban setting.

El Centro de Morfología Matemática (CMM), fundador de la teoría que lleva su nombre, es un laboratorio de investigación en tratamiento de imagen de la Escuela de Minas de Paris (Mines ParisTech). Siempre motivado por las aplicaciones industriales, el CMM ha desarrollado herramientas muy eficaces de tratamiento de imagen como por ejemplo los filtros conexos, que eliminan el ruido de la imagen sin afectar los contornos de los objetos de interés o la segmentación jerárquica, que combinada con la teoría de grafos da lugar a algoritmos muy eficientes, tanto en tiempo de cálculo como en calidad de resultados. El CMM ha participado con éxito en muchos proyectos de investigación en diversos campos (biología, medicina, multimedia, materiales, ...). Un rápido panorama de los proyectos en curso actualmente será presentado. Se presentará con detalle los resultados de dos proyectos recientes:

• TerraNumerica, en el que hemos desarrollado herramientas de análisis de escenas urbanas. El trabajo se centra en el análisis de nubes de puntos de escenas urbanas. El objetivo es extraer información semántica para una modelización realista. En primer lugar se segmenta la imagen en fachada, suelo u objeto presente en la escena. A continuación se clasifican los diferentes objetos de la escena (coche, peaton, farola...).

• I-towns, en el que hemos puesto a punto un método de localización de texto en imágenes genéricas, que permite extraer información semántica de las imágenes para realizar búsquedas de alto nivel.

A pentapod is usually defined as a 5-degree-of-freedom fully-parallel manipulator with an axial spindle as moving platform. This kind of manipulators have revealed as an interesting alternative to serial robots handling axisymmetric tools. Their particular geometry permits that, in one tool axis, inclination angles of up to 90 degrees are possible thus overcoming the orientation limits of the classical Stewart platform. This talk presents a solution to the problem of finding those changes in the location of the leg attachments of a pentapod that leave its singularity locus invariant. Although the solution to this problem does not provide a fully characterization of the singularities, it provides a lot of insight into its nature. It is shown, for example, that there are four different architectures for a pentapod with a completely different behavior from the point of view of their singularities. The kinematics of pentaponds with coplanar attachments at the fixed base has previously been studied as rigid subassemblies of a Stewart platforms. In this talk, we treat the general case in which the base attachments are arbitrarily located in 3D space.

Fourty years ago, Prof. L.O. Chua postulated the existence of a fourth element in the circuit theory; the Memristor. This element should complete, together with the conventional elements (resistors, inductances and capacitors), all the relationships between the electrical variables (currents, voltages, fluxes and charges). But it only existed on paper in the circuit theory... until 2008, when HP researchers found this missing element. This presentation is a big picture of the state of the art on the Memristor element. It starts with the definition and properties of the memristor and some possibles applications are pointed out. The universe of the two-terminal higher order elements is also introduced, and some of these elements are illustrated through easy examples. Finally, the port-Hamiltonian formalism is proposed to model these kind of elements. As conclusions, the open questions that arise during our research are presented.

This talk will cover the development of several autonomous robots with applications in education and entertainment and a unique robot theater project in the Advanced Intelligent Robotics Lab, Taiwan Tech (National Taiwan University of Science and Technology). The autonomous robots include desktop two-armed companion robots, DOC-1/DOC-2/DOC-3, which interactive functions in language teaching, board games, etc. The robot theater project developed four large sized humanoid robots, two two-wheeled, and two bipedal. The two-wheeled robots with two-arms can balance and move with two wheels, and can perform shows such as real-time portrait and music note reading-and-singing. The two bipedal robots are androids with two-arms and a human-like face capable of making facial expressions. They can walk and dance. The four robots together performed 5 shows to public in a robot theater in 27th December 2008. A new human-robot platform will be introduced at the end, which can significantly speed up the process and time for asking robots to operate on objects as a new assignment.

In this talk I will give a brief overview of the activities in the Imaga Analysis Lab in Stony Brook University. I will focus on our work in 3D Facial Analysis and present our most recent techniques for non-rigid surface matching. I will describe a high-order graph matching formulation to address non-rigid surface matching. The singleton terms capture the geometric and appearance similar- ities (e.g., curvature and texture) while the high-order terms model the intrinsic embedding energy. The novelty of this paper includes: 1) casting 3D surface registration into a graph matching problem that combines both geometric and appearance similarities and intrinsic embedding informa- tion, 2) the first implementation of high-order graph matching algorithm that solves a non-convex optimization prob- lem, and 3) an efficient two-stage optimization approach to constrain the search space for dense surface registration. Our method is validated through a series of experiments demonstrating its accuracy and efficiency, notably in challenging cases of large and/or non-isometric deformations, or meshes that are partially occluded.

We propose an efficient method, built on the popular Bag of Features approach, that obtains robust multiclass pixellevel object segmentation of an image in less than 500ms, with results comparable or better than most state of the art methods. We introduce the Integral Linear Classifier (ILC),that can readily obtain the classification score for any image sub-window with only 6 additions and 1 product by fusing the accumulation and classification steps in a single operation. In order to design a method as efficient as possible, our building blocks are carefully selected from the quickest in the state of the art. More precisely, we evaluate the performance of three popular local descriptors, that can be very efficiently computed using integral images, and two fast quantization methods: the Hierarchical K-Means, and the Extremely Randomized Forest. Finally, we explore the utility of adding spatial bins to the Bag of Features histograms and that of cascade classifiers to improve the obtained segmentation. Our method is compared to the state of the art in the difficult Graz-02 and PASCAL 2007 Segmentation Challenge datasets.

The standard forward kinematics analysis of 3-RPR planar parallel platforms boils down to computing the roots of a sextic polynomial. There are many different ways to obtain this polynomial but all of them include exceptions for which the formulation is not valid. Unfortunately, near these exceptions the corresponding polynomial exhibits numerical instabilities. In this paper, we provide a way around this inconvenience by translating the forward kinematics problem to be solved into an equivalent problem fully stated in terms of distances. Using constructive geometric arguments, an alternative sextic -which is not linked to a particular reference frame- is straightforwardly obtained without the need of variable eliminations nor tangent-half-angle substitutions. The presented formulation is valid, without any modification, for any planar 3-RPR parallel platform, including the special architectures and configurations -which ultimately lead to numerical instabilities- that cannot be directly handled by previous formulations.

This paper introduces a new method for workspace boundary determination on general lower-pair multi-body systems. The method uses a branch-and-prune technique to isolate the set of end effector singularities, and then classifies the points in such set according to whether they correspond to actual motion impediments in the workspace. The method can deal with open- or closed-chain systems, and is able to take joint limits into account. Advantages over other methods of similar applicability include its completeness and a simpler algorithmic structure. Examples are included that show its performance on benchmark problems documented in the literature.

In this talk I will describe the research activities at the Intelligent Systems Lab of the Institute for Systems and Robotics at Instituto Superior Técnico, Technical University of Lisbon. Most of the talk will concentrate on research done on networked robot systems, composed of static and mobile sensors that cooperate to carry out a diversified set of tasks. The group is particularly interested on discrete event models of robotic tasks, prone to qualitative (e.g., formal verification) and quantitative (e.g., plan probability of success) analysis, as well as on active cooperative perception and planning under uncertainty. A brief overview of past and running projects on soccer robots, rescue robots and mobile robots cooperating with a camera network to interact with people will be presented.

This work presents a necessary and sufficient condition to define a singularity-invariant leg rearrangement, based on an affine relation between the squared leg lengths before and after the rearrangement. This condition is then specified for four rigid components that can occur in Stewart-Gough platforms, leading to the characterization of singularity-invariant leg rearrangements on all of them.

Open cathode, self humidified PEM fuel cell (OCSH-PEMFC) systems have an intrinsic* *advantage over conventional systems due to their minimal peripheral components and high system power densities, however they tend to be affected by environmental conditions, operating modes and load changes. The main reason for this sensitivity is water distribution. Thus, fundamental understanding of water distribution and transport is imperative in this highly dynamic system, which can be strongly affected by different control strategies. In this context, a two-dimensional, non-isothermal, dynamic model of an OCSH-PEMFC was developed and used in conjunction with a custom built test station and environmental chamber. With these tools a comprehensive analysis of the system performance under a wide range of environmental and operating conditions associated with their internal physical phenomena is presented. This model is used to simulate and study the effects of water transport/distribution and their influence on the system performance in order to develop new water management control strategies to improve efficiency and operating range of a commercially available OCSH-PEMFC system. The model consists of three sub-models based on energy, momentum and water mass balances of the system applied to anode, cathode channel, diffusion layers and MEA. The model is also applicable to other PEMFC systems, following the developed modeling strategy and performing the proposed experiments in order to determine the particular coefficients.

We present a new approach for building an efficient and robust classifier for the two class problem, that localizes objects that may appear in the image under different orientations. In contrast to other works that adddress this problem using multiple classifiers, each one specialized for a specific orientation, or using multi-class classifiers, we propose a more straightforward approach that consists of two simple stages, namely estimation and classification. The estimator yields an initial object orientation hypothesis, which is then used by the classifier. False positive orientations provided by the estimator are rejected during the classification stage. This methodology allows reducing the time complexity of the algorithm while classification results remain high. The classifier we use in both stages is based on a boosted combination of Random Ferns over local histograms of oriented gradients (HOGs) in order to capture local statistics from gradient-based features. These are also another remarkable differences with respect to current methods where Random Ferns are computed over image intensity domain and without a supervised learning. We evaluate our method on standard databases, and show that it yields competitive results comparable to state-of-the-art approaches while being significantly more efficient. Furthermore, a new database for object detection under in-plane rotations is presented. This dataset contains motorbike instances with challenging conditions such as cluttered background, different illumination conditions and partial occlusions.

The homography between pairs of images are typically computed from the correspondence of keypoints, which are established by using image descriptors. When these descriptors are not reliable, either because of repetitive patterns or large amounts of clutter, additional priors need to be considered. The Blind PnP algorithm makes use of geometric priors to guide the search for matches while computing camera pose. Inspired by this, we propose a novel approach for homography estimation that combines geometric priors with appearance priors of ambiguous descriptors. More specifically, for each point we retain its best candidates according to appearance. We then prune the set of potential matches by iteratively shrinking the regions of the image that are consistent with the geometric prior. We can then successfully compute homographies between pairs of images containing highly repetitive patterns and even under oblique viewing conditions.

Recently, error minimized extreme learning machines (EM-ELMs) have been proposed as a simple and efficient approach to build single-hidden-layer feedforward networks (SLFNs) sequentially. They add random hidden nodes one by one (or group by group) and update the output weights incrementally to minimize the sum-of-squares error in the training set. Other very similar methods that also construct SLFNs sequentially had been reported earlier with the main difference that their hidden-layer weights are a subset of the data instead of being random. By analogy with the concept of support vectors original of support vector machines (SVMs), these approaches can be referred to as support vector sequential feedforward neural networks (SV-SFNNs). An experimental study on ten benchmark classification data sets, comparing EM-ELMs and SV-SFNNs, was carried out under the same conditions for the two models. Although both models have the same (efficient) computational cost, a statistically significant improvement in generalization performance of SV-SFNNs vs. EM-ELMs was found in six out of the ten benchmark problems.

In parallel mechanisms, singular configurations (singularities) have to be avoided during motion. All the singularities should be located in order to avoid them. Hence, relationships involving all the singular platform poses (singularity locus) and the mechanism geometric parameters are useful in the design of parallel mechanisms. A "simple" expression of the singularity condition of the most general mechanism (6-6 FPM) of a class of parallel mechanisms usually named fully-parallel mechanisms (FPM) is deduced by using a particular Laplace expansion of its Jacobian's determinant. The presented expression uses the mixed products of vectors that are easy to be identified on the mechanism. This approach will permit some singularities to be geometrically found. A procedure, based on this expression, is provided to transform the singularity condition into a ninth-degree polynomial equation whose unknowns are the platform pose parameters. This singularity polynomial equation is cubic in the platform position parameters and a sixth-degree one in the platform orientation parameters. Finally, how to derive the expression of the singularity condition of a specific FPM from the presented 6-6 FPM singularity condition will be shown along with an example.

PEM fuel cells are considered to have the highest energy density due to the nature of the reaction and the quickest start up time without pollution. These are the main reasons for being used in applications such as automotive engines, portable and backup power applications. Recent years have seen a proliferation of PEM FC system optimization and control applications, where the aim is to obtain a better process performance. This presentation addresses the problem of a model based fault diagnosis methodology for a PEM fuel cell system as case study. The methodology is based on computing residuals, which are obtained by PEM FC system model which has been calibrated using real data. The innovation of this methodology is based on the characterization of the relative residual fault sensitivity. To illustrate the results, a non-linear fuel cell simulator is used.

A 5-SPU robot with collinear universal joints is well suited to handling an axisymmetric tool, since it has 5 controllable DoFs and the remaining one is a free rotation around the tool. The kinematics of such a robot having also coplanar spherical joints has previously been studied as a rigid subassembly of a Stewart-Gough platform, it being denoted a line-plane component. It was shown that this component has 8 assembly modes corresponding to the roots of a bi-quartic polynomial. Here we identify a whole family of these 5-SPU robots having only 4 assembly modes, which are obtained by solving two quadratic equations. This family is defined by a simple proportionality constraint relating the coordinates of the base and platform attachments. A geometric interpretation of the architectural singularities of this type of robots in terms of conics is provided, which facilitates their avoidance at the design stage. Parallel singularities obey also a neat geometric structure, which permits deriving a cell decomposition of configuration space. Two practical features of these quadratically-solvable robots are the large maneuverability within each connected component and the fact that, for a fixed orientation of the tool, the singularity locus reduces to a plane.

This paper introduces a class of parallel robots consisting of a fixed base and a moving platform connected by serial chains having RR_PS (Revolute-Revolute-Prismatic-Spherical) topology. Only the prismatic joint is actuated and the first revolute joint in the chain can be locked or released online. The introduction of these lockable joints allow the prismatic actuators to manoeuver to approximate 6-DoF motions for the moving platform. An algorithm for generating these maneuvers is first described. Then, a motion planner, based on the generation of a Probabilistic RoadMap (PRM) whose nodes are connected using the presented maneuvers, is presented. The generated trajectories avoid singularities and possible collisions between legs.

Time-of-Flight (ToF) cameras deliver 3D images at 25 fps, offering great potential for developing fast object modeling algorithms. Surprisingly, this potential has not been extensively exploited up to now. A reason for this is that, since the acquired depth images are noisy, most of the available registration algorithms are hardly applicable. A further difficulty is that the transformations between views are in general not accurately known, a circumstance that multi-view object modeling algorithms do not handle properly under noisy conditions. In this work, we take into account both uncertainty sources (in images and camera poses) to generate spatially consistent 3D object models fusing multiple views with a probabilistic approach. We propose a method to compute the covariance of the registration process, and apply an iterative state estimation method to build object models under noisy conditions.

The project aim is based on the undeniable fact that nowadays humanity development is dramatically linked to the energy consumption. Our social and industrial way of life wastes in minutes what has need millions to be generated and creates residues that become hazardous thousands of years. But even thought, there is a bigger problem, there are not enought reserves that guarantee our growth level for the next half of the century. As a result we focus our end of degree project on Proton Exchange Membrane (PEM). PEM uses hydrogen as energetic vector to produce electricity by a simple catalytic process. To test that capabilities we will try to face to face our PEM prototype with a self transportation device with its own power supply system. This device target will be a Segway RMP-200 and more specifically the NiMH power supply batteries on the package.

The Project Step by Step:

• Segway power supply characterization.

• Modeling operating load profiles.

• Pre-design of the power supply system using hydrogen fuel cells.

• Implementation of the Power supply prototype.

• Lab testing of the Prototype.

• Final design.

• On board Testing.

Learning action rules on-line, from an empty rule base, and using them immediately for decision making involve many difficulties. On-line learning demands techniques that rapidly adjust their estimation with new incoming examples. Additionally, if decision making is to occur intertwined with learning, learned rules should become available as soon as possible to assure an improvement in performance, and to achieve the goal in a reasonable amount of time. For this, rules should be learned from few experiences of actions relevant for the task, and decision making should not be time consuming.

Most machine learning techniques are designed to work in batch mode, i.e. when the samples for learning are available in advance, and use an iterative process that requires a significant amount of data and computation. The mentioned requirements for rule learning and decision making pose obstacles for the existing rule learning paradigms, where on-line learning is not considered, a significant amount of prior knowledge has to be provided, or a large number of experiences are required.

This talk proposes a method for on-line learning of action rules from few observations using a rule performance evaluation based on densities. It is also presented a method for macro rules generation using a new technique of condition propagation that avoids large computational cost to make decisions. Macro rules could significantly reduce the amount of deliberation as they might merge repetitive sequences of actions, or plans found with a large computational cost. Other approaches have focused on the generation of macros such as primitive behaviors, macro-actions, or activation rules. They explore different long-term acting behaviors to select the one suitable for the task which requires a large amount of computation.

The learning mechanisms are embedded in a Decision Making Framework (DMF) that closes effectively the planning-learning loop. The DMF integrates: the rule learning methods, a planner, and an expert teacher that guides action executions.

See the associated document.

See the associated document.

This Master's thesis has two main contributions. Firstly, reviews the state-of-the art in the field of Time-of-Flight (ToF) cameras, their advantages, their limitations, the existing calibration methods and their present-day applications sometimes in combination with other sensors. And secondly, proposes a method for object modelling using a ToF camera under an uncertainty reduction approach.
Even though ToF cameras provide neither higher resolution nor larger ambiguity-free range compared to other range map estimation systems, advantages such as registered depth and intensity data at a high frame rate, compact design, low weight and reduced power consumption have motivated their use in numerous areas of research. In robotics, these areas range from mobile robot navigation and map building to vision-based human motion capture and gesture recognition, showing particularly a great potential in object modelling and recognition. ToF cameras deliver 3D images at 25 fps, offering great potential for developing fast object modelling algorithms. Surprisingly, this potential has not been extensively exploited up to now. A reason for this is that, since the acquired depth images are noisy, most of the available registration algorithms are hardly applicable. A further difficulty is that the transformations between views are in general not accurately known, a circumstance that multiview object modelling algorithms do not handle properly, especially under noisy conditions. In this work, both uncertainty sources (in images and camera poses) to generate spatially consistent 3D object models fusing multiple views with a probabilistic approach are taken into account. A method to compute the covariance of the registration process is proposed, and a statistical framework applied to build object models under noisy conditions.

Many engineering and scientific problems can be formulated as a pattern recognition one. In such problems, linear methods are preferred for their simplicity and tractability. Unfortunately, linear methods have many limitations, of which several are still unknown. Understanding these limitations is key to advancing the current state of the art. In this talk, we will address these issues within the context of feature extraction and classification. We will define when linear feature extraction methods do not work and how this knowledge can be used to propose algorithms that are guaranteed to work in a large number of applications. Several results in vision, linguistics, and the modeling of the primary visual cortex will be presented. Time permitting, we will sketch the problem posed by classical normalization procedures. In particular, that of norm normalization, generally used to make texture invariant to the intensity of the light source, shape to scale and rotation, and for modeling mtDNA in genetics. This result will lead us to the definition on of new type of kernels -- Rotation Invariant Kernels -- which are specially useful in shape analysis problems. Open problems will be outlined during the course of the talk.

We present an approach to the problem of 3D map building in urban settings for service robots, using three dimensional laser range scans as the main sensor data input. Our system is based on the probabilistic alignment of 3D point clouds employing a delayed-state information-form SLAM algorithm, for which we can add observations of relative robot displacement efficiently. These observations come from the alignment of dense range data point clouds computed with a variant of the iterative closest point algorithm. The datasets were acquired with our custom built 3D range scanner integrated into a mobile robot platform. Our mapping results are compared to a GIS-based CAD model of the experimental site. The results show that our approach to 3D mapping performs with sufficient accuracy to derive traversability maps that allow our service robots navigate and accomplish their assigned tasks on a urban pedestrian area.

En esta charla, Vicenç hablará acerca de diversos trabajo desarrollados (o en desarrollo) dentro de las actividades del grupo SAC (Sistemes Avançats de Control), todos ellos relacionados con el modelado y control de redes de agua potable, redes de alcantarillado y redes de canales y riego. Adicionalmente, se comentarán diversos matices en cuanto a investigaciones no sólo ligadas con el modelado dinámico y gestión de tales redes, sino también temas de calidad del agua, propagación de cloro, eficiencia de gestión de redes, entre otros tópicos.

Artificial intelligence research is ushering in a new era of sophisticated, mass-market transportation technology. While computers can already fly a passenger jet better than a trained human pilot, people are still faced with the dangerous yet tedious task of driving automobiles. Intelligent Transportation Systems (ITS) is the field of study that aims to use artificial intelligence to make transportation safer, cheaper, and more efficient. Recent advances in ITS point to a future in which vehicles themselves handle the vast majority of the driving task. Once autonomous vehicles become popular, autonomous interactions amongst *multiple* vehicles will be possible. Current methods of vehicle coordination, which are all designed to work with human drivers, will be outdated. The bottleneck for roadway efficiency will no longer be the drivers, but rather the mechanism by which those drivers' actions are coordinated. While open-road driving is a well-studied and more-or-less-solved problem, urban traffic scenarios, especially intersections, are much more challenging.
This talk will address the question: To what extent and how can a multiagent intersection control mechanism take advantage of the capabilities of autonomous vehicles in order to make automobile travel safer and faster? First, I will introduce and specify the problem of intersection management as a multiagent system and define a metric by which solutions can be evaluated. Next, I will propose a novel multiagent intersection control mechanism in which autonomous driver agents call ahead and reserve space-time in the intersection, pending the approval of an arbiter agent called an intersection manager, which is located at the intersection.
Experiments indicate that our method can reduce delays at intersections by up to two orders of magnitude.

Emergencies in industrial warehouses are a major concern for fire fighters. The large dimensions together with the development of dense smoke that drastically reduces visibility, represent major challenges.
The Guardians robot swarm is designed to assist fire fighters in searching a large warehouse. In this talk I discuss the technology developed for a swarm of robots assisting fire fighters. I will briefly explain the swarming algorithms which provide the functionality by which the robots react to and follow humans. Next I discuss the wireless communication system, which is a so-called mobile ad-hoc network. The communication network provides also the means to locate the robots and humans. Thus the robot swarm is able to provide guidance information to the humans. Together with the fire fighters the team explored how the robot swarm should feed information back to the human fire fighter. We have designed and experimented with interfaces for presenting swarm based information to human beings.
To conclude: the interaction from the human to the robots is very distinct from the interaction of the robots with the human.

1. Simply by moving around a human being provokes reactions from the autonomous robots; the swarming algorithms provide this functionality. Thus the human to robots swarm interface requiring very little cognitive effort.

2. The robots can transfer guidance information to the human. In collaboration with the fire fighters several interfaces for obtaining guidance from the surrounding robot swarm have been designed and tested. Nevertheless, it is an outstanding issue whether a feel of confidence can be created.

Not disclosed.

This paper presents an asynchronous particle filter algorithm for mobile robot position tracking, taking into account time considerations when integrating observations being delayed or advanced from the prior estiamate time point. The interest of that filter lies in cooperative environments and in fast vehicles. The paper studies the first case, where a sensor network shares perception data with running robots that receive accurate obeservations with large delays due to acquisition, processing and wireless communications. Promising simulated results comparing a basic particle filter and the proposed one are shown. The paper also investigates a situation where a robot is tracking its position, fusing only odometry and observations from a camera network partially covering the robot path.

Outdoor camera networks are becoming ubiquitous in critical urban areas of large cities around the world. Although current applications of camera networks are mostly limited to video surveillance, recent research projects are exploiting advances on outdoor robotics technology to develop systems that put together networks of cameras and mobile robots in people assisting tasks. Such systems require the creation of robot navigation systems in urban areas with a precise calibration of the distributed camera network. Despite camera calibration has been an extensively studied topic, the calibration (intrinsic and extrinsic) of large outdoor camera networks with no overlapping view fields, and likely to suffer frequent recalibration, poses novel challenges in the development of practical methods for user-assisted calibration that minimize intervention times and maximize precision. In this paper we propose the utilization of Laser Range Finder (LRF) data covering the area of the camera network to support the calibration process and develop a semi-automated methodology allowing quick and precise calibration of large camera networks. The proposed methods have been tested in a real urban environment and have been applied to create direct mappings (homographies) between image coordinates and world points in the ground plane (walking areas) to support person and robot detection and localization algorithms.

Hydrogen fed polymer electrolyte membrane fuel cells (PEMFC) are energy conversion devices that can be implemented into a wide range of applications such as automotive, stationary and portable systems. However, to optimize performance, they require active control and thus in-depth understanding of the system dynamics. Understanding water transport mechanisms through the membrane and membrane water content is one of the keys to improving PEMFC system performance. Once these mechanisms are identified and modelled, effective control strategies can be implemented. Hence, the main objective of this test station design is to experimentally characterize the water transport mechanisms through the membrane and its water content for the purpose of studying their dynamic influence on fuel cell system performance for numerical model verification and validation. The analysis will then lead to an in-depth understanding of how a system with minimal parts and complexity can be improved in terms of overall efficiency, stability and operating rage. These conclusions will provide guidelines to develop a proper control strategy with the necessary sensors to ensure that the system operates properly under a wide range of environmental conditions.

We present a new model for people guidance in urban settings using one or several mobile robots, that overcomes the limitations of existing approaches, which are either tailored to tightly bounded environments, or based on unrealistic human behaviors. Although the robots motion is controlled by means of a standard particle filter formulation, the novelty of our approach resides in how the environment and human and robot motions are modelled. In particular we define a Discrete-Time-Motion model, which from one side represents the environment by means of a potential field, that makes it appropriate to deal with open areas, and on the other hand the motion models for people and robots respond to realistic situations, and for instance human behaviors such as leaving the group are considered.

Time to Contact (TTC) is a biologically inspired method for obstacle detection and reactive control of motion that does not require scene reconstruction or 3D depth estimation. Estimating TTC is difficult because it requires a stable and reliable estimate of the rate of change of distance between image features. In this paper we propose a new method to measure time to contact, Active Contour Affine Scale (ACAS). We experimentally and analytically compare ACAS with two other recently proposed methods: Scale Invariant Ridge Segments (SIRS), and Image Brightness Derivatives (IBD). Our results show that ACAS provides a more accurate estimation of TTC when the image flow may be approximated by an affine transformation, while SIRS provides an estimate that is generally valid, but may not always be as accurate as ACAS, and IBD systematically over-estimate time to contact.

Time to Contact (TTC) is a biologically inspired method for obstacle detection and reactive control of motion that does not require scene reconstruction or 3D depth estimation. TTC is a measure of distance expressed in time units. Our results show that TTC can be used to provide reactive obstacle avoidance for local navigation. In this paper we describe the principles of time to contact and show how time to contact can be measured from the rate of change of size of features. We show an algorithm for steering a vehicle using TTC to avoid obstacles while approaching a goal. We present the results of experiments for obstacle avoidance using TTC in static and dynamic environments.

The computational bottleneck in all information-based algorithms for SLAM is the recovery of the state mean and covariance. The mean is needed to evaluate model Jacobians and the covariance is needed to generate data association hypotheses. Recovering the state mean and covariance requires the inversion of a matrix of the size of the state. Current state recovery methods use sparse linear algebra tools that have quadratic cost, either in memory or in time. In this paper, we present an approach to state estimation that is worst case linear both in execution time and in memory footprint at loop closure, and constant otherwise. The approach relies on a state representation that combines the Kalman and the information-based state representations. The strategy is valid for any SLAM system that maintains constraints between robot poses at different time slices. This includes both Pose SLAM, the variant of SLAM where only the robot trajectory is estimated, and hierarchical techniques in which submaps are registered with a network of relative geometric constraints.

This paper presents an efficient graph-based algorithm for the segmentation of planar regions out of 3D range maps of urban areas. Segmentation of planar surfaces in urban scenarios is challenging because the data acquired is typically sparsely sampled, incomplete, and noisy. The algorithm is motivated by Felzenszwalb's algorithm to 2D image segmentation , and is extended to deal with non-uniformly sampled 3D range data using an approximate nearest neighbor search. Inter-point distances are sorted in increasing order and this list of distances is traversed growing planar regions that satisfy both local and global variation of distance and curvature. The algorithm runs in O(n log n) and compares favorably with other region growing mechanisms based on Expectation Maximization. Experiments carried out with real data acquired in an outdoor urban environment demonstrate that our approach is well-suited to segment planar surfaces from noisy 3D range data. A pair of applications of the segmented results are shown, a) to derive traversability maps, and b) to calibrate a camera network.

This thesis proposal addresses the problem of building maps for mobile robots that navigate in outdoor urban areas. And we wish to do so as autonomously as possible. In this context, our robot must be able to build a locally detailed and globally consistent map using 3D range sensing as the primary source of information. The robot must be able to determine its position with 6 degrees-of-freedom (DOF) within this map. This task is known in the robotics literature as Autonomous Simultaneous Localization and Mapping. Furthermore, from the acquired maps, the robot must be able to perform traversability analyses to take informed decissions for navigation and exploration. Once the map is built some post-processing might be necessary to obtain a complete digital compact volumetric model, merging information to reduce the memory space of the map without compromising details and consistency. The presentation includes a brief analysis of the state of the art in the topic, as well as some initial contributions on efficient algorithms for range data registration, map building in the context of the URUS project, and applications of the produced maps to traversability analysis and camera network calibration.

The talk deals with the problem of motion planning of anthropomorphic mechanical hands avoiding collisions. The approach that will be presented tries to mimic the real human hand motions, but reducing the dimension of the search space in order to obtain results as a compromise between motion optimality and planning complexity (time) by means of the concept of principal motion directions. Basically, the work includes the following phases: capturing the human hand workspace using a sensorized glove and mapping it to the mechanical hand workspace, reducing the space dimension by looking for the most relevant principal motion directions, and planning the hand movements using a sampling-based roadmap planner. The approach has been implemented for a four finger anthropomorphic mechanical hand, and some examples are will be shown to illustrate its validity.

Simultaneous Localization and Mapping (SLAM), Smoothing and Mapping (SAM), and Structure from Motion (SFM) are important and closely related problems in robotics and vision. Not surprisingly, there is a large literature describing solutions to each problem, and more and more connections are established between the two fields. At the same time, robotics and vision researchers alike are becoming increasingly familiar with the power of graphical models as a language in which to represent inference problems. In this talk I will show how SFM, SAM, and SLAM can be posed in terms of this graphical model language, and how inference in them can be explained in a purely graphical manner via the concept of variable elimination. I will then present a new way of looking at inference that is equivalent to the junction tree algorithm yet is - in my view- much more insightful. I will also show that, when applied to linear(ized) Gaussian problems, the algorithm yields the familiar QR and Cholesky factorization algorithms, and that this connection with linear algebra leads to strategies for very fast inference in arbitrary graphs. I will conclude by showing some published and preliminary work that exploits this connection to the fullest.

In this work we present a robust detection method in outdoor scenes under cast shadows using color based invariant gradients in combination with HoG local features. The method achieves good detection rates in urban scene classification and person detection outperforming traditional methods based on intensity gradient detectors which are sensible to illumination variations but not to cast shadows. The method uses color based invariant gradients that emphasize material changes and extract relevant and invariant features for detection while neglecting shadow contours. This method allows to train and detect objects and scenes independently of scene illumination, cast and self shadows. Moreover, it allows to do training in one shot, that is, when the robot visits the scene for the first time.

Pose SLAM is the variant of SLAM where only the robot trajectory is estimated and where landmarks are only used to produce relative constraints between robot poses. To reduce the computational cost of the information filter form of Pose SLAM and, at the same time, to delay inconsistency as much as possible, we introduce an approach that only takes into account highly informative loop closure links and non redundant poses. The technique includes constant time procedures to compute the distance between poses and the potential information gain obtained from each link, as well as an approximately linear procedure in terms of time and memory, to recover the state after a loop closure. Using these procedures, the robot operates most of the time in open loop, amortizing the cost of loop closure over long trajectories. In this case, the computational bottleneck shifts to data association, that is, to a search over the set of previously visited poses to determine good candidates for sensor registration. To speed up data association, we introduce a method to search for neighboring poses whose complexity ranges from logarithmic in the usual case to linear in degenerate situations. The method is based on organizing the pose information in a balanced tree whose internal levels are defined using interval arithmetic. The proposed method is validated through simulations, experiments using standard SLAM data sets, and real mapping sessions.

When the attachments of five legs in a Stewart platform are collinear on one side and coplanar on the other, the platform is said to contain a line-plane subassembly. This talk is devoted to the kinematics analysis of this subassembly paying particular attention to the problem of moving the aforementioned attachments without altering the singularity locus of the platform. It is shown how this is always possible provided that some cross-ratios between lines ---defined by points in the plane--- are kept equal to other cross-ratios between points in the line. This result leads to two simple motion rules upon which complex changes in the location of the attachments can be performed. These rules have interesting practical consequences as they permit a designer to optimize aspects of a parallel robot containing the analyzed subassembly, such as its manipulability in a given region, without altering its singularity locus.

In this presentation I will discuss a course I have developed at the University of Ottawa, Canada, that covers the ethical, legal and social issues important for computer professionals. I will describe the contents of the course that include: an introduction to computer ethics and how to apply it in practice, privacy in the electronic era, social implications of the Internet, computer failures and legal responsibility, professional codes of conduct, global issues including digital divide, and the ethical issues related to robotics and autonomous computing. I will describe my experience and the student experience from three years of teaching this course in Ottawa. I will discuss a sample textbook and exam material for this course. I will argue that such a course (already a compulsory component of the ACM/IEEE curriculum in all computer science and engineering disciplines) truly belongs, for academic and moral reasons, to any modern information technology program.

The brain is still the most advanced system for perception, cognition and control that we know. Unfortunately we still do not have a theory of the brain that would allow us to copy neuronal principles to artifacts. Over the last few years, however, a number of inroads have been made in the areas of perception, learning and motor control. In my talk I will focus on our own work within the framework of the Distributed Adaptive Control architecture and its implications for our understanding of the neuronal principles underlying real-world perceptual systems, mapping and localization and the planning of discrete motor actions and the generalization of these principles towards robots, avatars and neuroprosthetic systems.

Kinematic synthesis aims to solve the function-to-form problem when dealing with motion. It has its origins in the graphical design of planar mechanisms and it is now applied to spatial mechanisms and robotic systems. One of the branches of kinematic synthesis is dimensional synthesis, in which the structure (also called topology) of the system is given and the goal is to dimension it in order to accomplish a given kinematic task.
In this talk, I want to present some of the unsolved problems within the dimensional synthesis of kinematic chains for spatial motion. A brief introduction to different approaches and their limitations will be given, to then focus on the use of Clifford algebras to express the synthesis problem. Several applications in the field of robot design and motion identification will be presented.

In this talk I will describe a system that can carry out SLAM in large indoor and outdoor environments using a stereo pair moving with 6DOF as the only sensor. Unlike current visual SLAM systems that use either bearing-only monocular information or 3D stereo information, our system accommodates both monocular and stereo. Textured point features are extracted from the images and stored as 3D points if seen in both images with sufficient disparity, or stored as inverse depth points otherwise. This allows to map both near and far features: the first provide distance and orientation, and the second orientation information. Unlike other vision only SLAM systems, stereo does not suffer from scale drift because of unobservability problems, and thus no other information such as gyroscopes or accelerometers is required in our system. Our SLAM algorithm generates sequences of conditionally independent local maps that can share information related to the camera motion and common features being tracked. The system computes the full map using the novel Conditionally Independent Divide and Conquer algorithm, which allows constant time operation most of the time, with linear time updates to compute the full map. To demonstrate the robustness and scalability of our system, I will show experimental results in indoor and outdoor urban environments of 210m and 140m loop trajectories, with the stereo camera being carried in hand by a person walking at normal walking speeds of 4-5km/hour.

The coupling between form and forces, their structural morphology, is a key point for tensegrity systems. In the first part of this paper we describe the design process of the simplest tensegrity system which was achieved by Kenneth Snelson. Some other simple cells are presented and tensypolyhedra are defined as tensegrity systems which meet polyhedra geometry in a stable equilibrium state. A numerical model giving access to more complex systems, in terms of number of components and geometrical properties, is then evoked. The third part is devoted to linear assemblies of annular cells -tensegrity rings- which can be folded. Some experimental models of the tensegrity ring which is the basic component of this "hollow rope" have been realized and are examined. The last part is devoted to some possible foldable tensegrity systems.

Path planning is an active field of research in robotics since the 70s. The addressed problem consists of computing feasible motions for a system (the robot) in a workspace cluttered with obstacles. In recent years, path planning techniques have undergone considerable progress. Sampling-based algorithms are efficient and general tools for exploring constrained high-dimensional spaces. Such algorithms have been successfully applied to challenging problems in diverse domains, including computational structural biology. This seminar gives an overview of recent works carried out at LAAS-CNRS in this domain. In particular, I will present path-planning-based algorithms for computing protein loop/domain motions and protein-ligand access/exit pathways. I will also show encouraging results obtained with these methods on real application examples.

I will present a non-iterative solution to the PnP problem --the estimation of the pose of a calibrated camera from n 3D-to-2D point correspondences-- whose computational complexity grows linearly with n. This is in contrast to state-of-the-art methods that are O(n^5) or even O(n^8), without being more accurate. Our method is applicable for all n>=4 and handles properly both planar and non-planar configurations. Our central idea is to express the n 3D points as a weighted sum of four virtual control points. The problem then reduces to estimating the coordinates of these control points in the camera referential, which can be done in O(n) time by expressing these coordinates as weighted sum of the eigenvectors of a 12x12 matrix and solving a small constant number of quadratic equations to pick the right weights. Furthermore, if maximal precision is required, the output of the closed-form solution can be used to initialize a Gauss-Newton scheme, which improves accuracy with negligible amount of additional time. The advantages of our method are demonstrated by thorough testing on both synthetic and real-data.

Graphs have very interesting properties for object representation in pattern recognition. However, graph matching algorithms are usually computationally complex. In addition, graphs are harder to manipulate and operate than feature vectors. In the last years, some attempts have been made to combine the best of the graph and the vector domains in order to get the advantages of both worlds. In this sense, kernels offer an elegant way to embed graphs into vector spaces. In this work, we show how graph embedding can be used to obtain accurate and efficient approximations of the median graph. The median graph can be seen as the representative of a set of graphs but its application has been very limited up to now due to computational reasons. With this new approach, we can obtain an approximate median graph using real databases containing large graphs.

We present a method called Transition-based RRT (T-RRT) for path planning in continuous cost spaces. It combines the exploration strength of the RRT algorithm that rapidly grow random trees toward unexplored regions of the space, with the efficiency of stochastic optimization methods that use transition tests to accept or to reject a new potential state. This planner also relies on the notion of minimal work path that gives a quantitative way to compare path costs. The method integrates self tuning ofa parameter controlling its exploratory behavior. It yields to solution paths that efficiently follow low cost valleys and saddle points of the cost space. Simulation results show that the method can be applied to a large set of applications including terrain costmap motions or planning low cost motions for freeflying or articulated robots.

I will present recent advances in Computer Vision for Augmented Reality, and their commercial applications. I will also discuss the limits of the current techniques, and the possible directions to go beyond.

Online reinforcement learning with generalization is a problem still unsolved for many complex applications. The main obstacles are produced by the high non-stationarity of the estimated cumulative reward, the online nature of the learning with non-uniform sampling, and the large amount of generalization required for a feasible implementation. Our previous approaches to face the RL problem consisted in estimating the probability distribution of the cumulative reward in coarse discrete partitions of the state space using a rule-based representation. The discrete partition and the biased sampling inherent to RL produced bad estimations of the cumulative reward due to the limited information about the real distribution of data that fed the rules. To overcome this inconvenient we propose a new RL approach based on a joint distribution estimation of the samples (state, action, cumulative reward) experienced. From this joint distribution it is possible to estimate the distribution of the cumulative reward with the highest confidence possible given the data experienced so far. The joint distribution is modelled with a Gaussian Mixture estimated with the online expectation-maximization algorithm and with active unit manipulation.

Interval constraint solving techniques aim at returning an as-precise-as-desired framing of all the solutions to a system of equations and inequalities over the real numbers. They derive from numerical analysis and artificial intelligence: the numerical computations are performed on intervals instead of real numbers to avoid the inaccuracy of floating-point numbers computations and to enable a complete exploration of the search space; smart filtering techniques are employed in order to avoid searching the solutions in inappropriate regions of the search space. This talk will introduce these techniques and present their application to some robotics problems studied in the MEO group of the LINA lab to which the speaker belongs.

Selecting a good representation of the data is a key aspect of the success of any modeling, classification or clustering algorithm. Component Analysis (CA) methods (e.g. Kernel Principal Component Analysis, Independent Component Analysis, Tensor factorization) have been used as a feature extraction step for modeling, classification and clustering in numerous visual, graphics and signal processing tasks over the last four decades. CA techniques are especially appealing because many can be formulated as eigen-problems, offering great potential for efficient learning of linear and non-linear representations of the data without local minima. However, the eigen-formulation often hides important aspects of making the learning successful such as understanding normalization factors, how to build invariant representations (e.g. to geometric transformation), effects of noise and missing data or how to learn the kernel. In this talk, I will describe a unified framework for energy-based learning in CA methods. I will point out how apparently different learning tasks (clustering, classification, modeling) collapse into a single task when viewed from the perspective of energy functions. Moreover, I will propose several extensions of CA methods to learn linear and non-linear representations of data to improve performance, over the current use of CA features, in state-of-the-art algorithms for classification (e.g. support vector machines), clustering (e.g. spectral graph methods) and modeling/visual tracking (e.g. active appearance models) problems.
During the talk I will also review applications of CA techniques to depression assestment, deception detection and hot-flash prediction from multimodal sensors.

A brief summary of the history of tensegrity structures opens this talk, which summarizes the spaker's personal perspective on tensegrity structures. The technical part of the presentation focuses on two topics which have been investigated by the speaker in the past: tensegrity structures deployment and their use as motion base for flight simulators as replacements for the classical Stewart platform. The talk is concluded with considerations on tensegrity structures future and the challenges they face, especially in the field of controllable structures applications.

This talk gives a brief account of the activities developed during the first year of the URUS project
The talk was previously given in the ICRA'08 workshop Network Robot Systems: benchmarks and platforms toward Human-Robot Interaction, co-organized by Alberto Sanfeliu.

Flagged manipulators are of interest because they are the only Stewart-Gough platforms for which a cell decomposition of their singularity loci is available. Here we show that the known family of such manipulators can be enlarged if one allows robot designs that, for some particular parameter values, become architecturally singular. Along this line, the most general 6-6 flagged manipulator is derived by applying a singularity-preserving transformation that leaves the relative position between two lines invariant. This transformation opens up the possibility of an "equal cross ratios" architectural singularity, which is shown to appear clearly in the factorization of the jacobian determinant. From the 6-6 flagged manipulator, all the extended family of (possibly architecturally-singular) flagged manipulators is derived.

This paper presents a novel and efficient framework to the active map-based global localization problem for mobile robots operating in large and cooperative environments. The paper proposes a rational criteria to select the action that minimizes the expected number of remaining position hypotheses, for the single robot case and for the cooperative case, where the lost robot takes advantage of observations coming from a sensor network deployed on the environment or from other localized robots. Efficiency in time complexity is achieved thanks to reasoning in terms of the number of hypotheses instead of in terms of the belief function. Simulation results in a real outdoor environment of 10.000m2 are presented validating the presented approach and showing different behaviors for the single robot case and for the cooperative one.

Manipulation of deformable objects constitutes a challenge and a huge field of opportunities for roboticists, due to the richness in behaviour and the extended presence of such objects in everyday life. Long-term and goal-directed interaction of a robot with such objects implies to provide some kind of model-based planning capabilities. Whereas manipulation planning of rigid objects, viewed as an offspring of Motion Planning, is a well-studied subject, works dealing about manipulation planning of deformable objects are spread over a wide range of apparently unconnected application areas. This survey constitutes a first attempt of bringing results together and to present a systematic approach to the subject. The interested researcher will find here how such objects are currently modelled, which basic planning tools are available, how application-dependent constraints are dealt with and considered in planning, as well as a lot of inspiring examples.

Planning a precision grasp for a robot hand is usually decomposed into two main steps. First, a set of contact points over the object surface must be determined, ensuring they allow a stable grasp. Second, the inverse kinematics of the robot hand must be solved to verify whether the contact points can actually be reached. Whereas the first problem has been largely solved in a general posing, the second one has only been tackled with local convergence methods. These methods only provide one solution to the problem, even if many are possible, and depending on the initial estimation they use, they may fail to converge, which results in grasp re-planning in situations where it could be avoided. This paper overcomes both issues by providing a complete method to solve the kinematics of human-like hands. The method is able to find all possible configurations that reach the specified contact points, even when positive-dimensional sets of such configurations are possible.

This talk presents a numerical method for the position analysis of multi-loop robots and molecules. The method is general (it can be applied to single or multiple interconnected loops of arbitrary topology) and complete (it isolates all solutions, even if they form positive-dimensional sets). Generality is achieved by reducing the problem to finding all embeddings of a set of points constrained by pairwise distances, which can be formulated as computing the roots of a system of Cayley-Menger determinants. Completeness is achieved by expressing these determinants in Bernstein form and using a numerical algorithm that exploits such form to bound all root locations at any desired precision. The method is readily parallelizable, and the current implementation can be run on single- or multi-processor machines. Experiments are included that show the method's performance on rigid loops, mobile loops, and multi-loop molecules. In all cases, complete maps including all possible configurations are obtained, thus allowing an exhaustive analysis and visualization of all pseudo-rotation paths between different conformations satisfying loop closure.

The main difficulty to attain fully autonomous robot navigation outdoors is the fast detection of reliable visual references, and their subsequent characterization as landmarks for immediate and unambiguous recognition. Aimed at speed, our strategy has been to track salient regions along image streams by just performing on-line pixel sampling. Persistent regions are considered good candidates for landmarks, which are then characterized by a set of subregions with given color and normalized shape. They are stored in a database for posterior recognition during the navigation process. Some experimental results showing landmark-based navigation of the legged robot Lauron III in an outdoor setting are provided.

Flagged parallel manipulators are a special kind of parallel robots with nice kinematics and singularity properties. This seminar will first do an overview of parallel robots and the kinematic problems associated, talking about several mathematical tools we use to model them. Next, the study of flagged robots and their nice properties will be presented, as well the stratification of their singularity loci and geometric methods to generate a family of manipulators with the same properties.

In this presentation I will discuss the problem of augmenting ordinary objects and environments with information and communication technologies. I will begin with a presentation of projected interaction in the PRIMA Augmented Meeting Environment at INRIA Rhones-Alpes. I will present video demonstrations in which ordinary surfaces are augmented with abilities for presentation and interaction using a steerable camera-projector pair. I will then address the looming problem of disruption created by the large-scale use of such technologies. I present a possible solution to this problem that relies on providing environments with an understanding of the activity and context of activities. I will briefly describe a layered architecture for building context aware services for augmented environments, finishing with an example of an automated video acquisition systems for context aware recording and transmission of meetings and lectures.

In this talk I will present a novel algorithm for the computation of corresponding image regions of stereo pairs and motion sequences. In this framework, local correspondences synchronize gray-level segmentation of the image via an n-d conjoint spin relaxation process, which is based on the method of superparamagnetic clustering of data. Spin interactions result in the formation of clusters of correlated spins, providing an automatic labeling of corresponding image regions. I will show results for both synthetic and real image sequences, and suggest potential applications of the algorithm.

This talk presents an efficient approach to outdoor visually augmented odometry. The technique computes relative pose constraints via a robust least squares minimisation of 3D point correspondences, which are in turn obtained from the matching of SIFT features over two consecutive image pairs. Pose constraints are then used to build a history of pose estimates with and incremental delayed-state information filter. The efficiency of the approach resides on the exact sparseness of the delayed-state information form used. We also propose a loop closure test in a SLAM context that checks both for closeness of means and for highly informative updates at the same time.

Us parlarà de tensegrities, què són, perqué s'aguanten a l'espai, els diferents punts de vista des d'els quals els podem analitzar per obtenir les seves propietats de rigidés i estabilitat: moviments dels nodes, forces que suporten els membres, energia de l'estructura. En aquest repàs veurem l'estat de l'art, les aplicacions existents, el que hem après d'aquestes estructures i el que ens falta per aprendre. Després revisarem el que estem fent el grup de persones que hi treballem, així com els objectius dels dos projectes que actualment hi ha en tensegrity.

The presented work focuses on mobile robot dynamic localization based on the angular measurements of the orientation, relative to the robot frame, of the straight lines between one of the robot points and artificial landmarks. The measurements are assumed to be done by means of a rotating laser scanner that detects the different landmarks. It is well-known that under robot static condition, geometric triangulation methods can be used to determine the robot pose from the measured angles.
However, when the robot moves the landmarks are detected at different vehicle configurations and therefore the geometric methods cannot be consistently used to determine the robot pose directly from the measurements. This problem, known in robotics as the dynamic positioning problem, has been widely studied in the past decades and solved by many authors using a pose-state Kalman filter. This algorithm fuses the robot positional odometry with each of the laser measurements to estimate the robot pose. In this work, an angular-state vector is considered for the Kalman filter in order to estimate the landmark angles. By using this algorithm, the evolution of each landmark angle between actual laser measurements is obtained, and then the triangulation methods can be consistently used at any time to calculate the robot pose.
The work presents computer simulations and real experiments using the robot SPHERIK-3x3, which has an omnidirectional kinematics (with 3 DOF) and is based on three directionally sliding wheels. The results obtained both in the simulations and the experiments, show that the developed method performs better in terms of positioning accuracy.

In this talk an active and cooperative strategy to solve the map-based global localization problem for autonomous mobile robots is presented. Starting from a completely lost robot, common strategies such as particle filtering or place recognition result in a finite number of position hypotheses, specially in cases of large and real environments. For a fully autonomous system, the robot has to decide moving somewhere in order to disambiguate this set of hypotheses. Usually, places within the robot sensor horizons can not disambiguate the set of hypotheses, therefore a technique searching beyond them is required. This work presents a probabilistic strategy that uses the map of the environment to evaluate candidate positions in an exploration area around the robot. The strategy is neither sensor dependent nor spatial representation dependent and is easily extensible to multi-robot and sensor network contexts. Finally, an implementation of the proposed strategy is discussed and some simulation results in a real environment of 10000 m2 are shown validating the presented approach.

Haptics devices have had a big evolution in the last ten years and the number of different applications in Robotics has increased hugely. This has caused many researchers to be concerned with the development of better tactile and force feedback devices for supporting all areas where this kind of machines could be useful. The goals of this seminar are to review the basic concepts about haptics, to show the different kinds of haptic devices and their main features. The Delta Haptic Device recently purchased by the institute will be introduced. On other hand, the most important applications in Robotics will be explained. Current projects about haptic rendering in virtual environments and haptic feedback for Teleoperation tasks will be discussed in depth.

This presentation will serve as an overview of the research activities conducted at the Robotics and Mechanisms Laboratory (RaM Lab) at the University of New Brunswick (Canada). Current research work mainly concentrates on two general areas: the study of parallel manipulators and the simulation of robotic systems.
In the area or parallel manipulators, studies are being conducted on kinematic and actuation redundancy in order to improve workspace characteristics (e.g., workspace size and quality, force capabilities). Additionally, studies involving the development of new kinematic calibration methods for reduced degree-of-freedom parallel manipulators are also under way. The new methods are being validated through a test platform based on the 3-PRS parallel architecture.
In terms of the computer simulation of robotic systems, the current work focuses on the development of a simulation environment for tethered underwater vehicles. More particularly, new methods for determining the separation distance and contact status between convex and/or concave objects are currently being extended for flexible slender objects (e.g., ROV tether). Contact dynamics models for such objects have also been developed and are being incorporated into realistic tethered vehicle simulators.

Symmetries in discrete constraint satisfaction problems have been explored and exploited in the last years, but symmetries in continuous constraint problems have not received the same attention. Here we focus on permutations of the variables consisting of one single cycle. We propose a procedure that takes advantage of these symmetries by interacting with a Branch-and-Prune algorithm without interfering with it. A key concept in this procedure are the classes of symmetric boxes formed by bisecting a n-dimensional cube at the same point in all dimensions at the same time. We analyze these classes and quantify them as a function of the cube dimensionality. Moreover, we propose a simple algorithm to generate the representatives of all these classes for any number of variables at very high rates. A problem example from the chemical field and a kinematics solver are used to show the performance of the approach in practice.

Robot egomotion can be estimated from an acquired video stream up to the scale of the scene. To remove this uncertainty (and obtain true egomotion), a distance within the scene needs to be known. If no a priori knowledge on the scene is assumed, the usual solution is to derive ``in some way" the initial distance from the camera to a target object. We propose a new, very simple way to obtain such a distance, when a zooming camera is available and there is a planar target in the scene. Similarly to ``two-grid calibration" algorithms, no estimation of the camera parameters is required, and no assumption on the optical axis stability between the different focal lengths is needed. Quite the reverse, the non stability of the optical axis between the different focal lengths is the key ingredient that enables to derive our depth estimate, by applying a result in projective geometry. Experiments carried out on a mobile robot platform show the promise of the approach.

Wire-based tracking devices are an affordable alternative to costly tracking devices. They consist of a fixed base and a platform, attached to the moving object, connected by six wires whose tension is maintained along the tracked trajectory. One important shortcoming of this kind of devices is that they are forced to operate in reduced workspaces so as to avoid singular configurations. Singularities can be eliminated by adding more wires but this causes more wire interferences, and a higher force exerted on the moving object by the measuring device itself. In this talk we show how, by introducing a rotating base, the number of wires can be reduced to three, and singularities can be avoided by using an active sensing strategy. This also permits reducing wire interference problems and the pulling force exerted by the device.

The Spaceward Foundation manages NASA's annual beam power challenge that will be taking place during the X Prize Cup in New Mexico. Being the first contest of NASA's X-Prize inspired Centennial Challenges Program the beam power challenge requires to design, build and operate a climber, a machine just powered by a beam of light that can drive up and down a tether ribbon, while carrying a payload.
Entries are rated according to the product of payload weight and climbing speed, normalized by the net climber weight. This metric is representative of the throughput of a space transportation system based on the Space Elevator.
Our prototype climber will be composed by a panel of photovoltaic power receptors that power a climb motor, all this assembled in a lightweight aluminium structure. The traction mechanism is formed by wheels that gripe the ribbon. The climber will need some power conditioning and control circuits. These circuits control the climber speed and adjust the voltage and current supplied by the solar cells to the traction motor and check the bottom and top of the ribbon.
A very important part is the beam source system. This source has to be concentrated and to be similar to the solar light, because the solar cells that we will be employing are designed for this type of radiation. Hence we will be using a Xenon bulb that matches sun light.
The climber design includes a super capacitor that will be storing part of the energy supplied by the solar panel. The stored energy will allow to keep climbing the ribbon in a steady and continuous manner, even without light for a few seconds. In addition we are planning to use a Fresnel lens to be able to better focus the beam of light at an increased distance from the Xenon source, leading to improved power supply by the solar array at higher altitudes of the climber.
At the moment the project is in its initial phase. We have got a prime sponsor and have started assembling the basic parts of the prototype.

In the first part of my talk I will briefly present our previous work on transferring human movements to humanoid robot movements, learning from observations using primitives, and on foveated vision and object recognition. In the second part I shall present our current work on building object representations by manipulation. I will focus on how to exploit a humanoid robot's motor capabilities to learn representations of objects without having any prior information about them. We will show that by taking control of the object the humanoid can resolve the problems arising in systems that rely on purely bottom-up attention and segmentation, which enables the robot to acquire appearance models suitable for recognition.

Tensegrity structures are pre-stressable truss-like systems which mantain their form due to an intrincate balance of forces between a disjoint set of rigid elements (bars) and a continuous set of tensile elements (cables). One of the fundamental problems of such structures is, given a description of the elements of the structure (cables or bars) and their topological relationship, find a stable configuration. This process implies both geometric and static constraints, and has received a lot of attention since this kind of structures were introduced. In this talk we propose a form-finding method based on non-linear programming which takes into account both geometric and static constraints and is shown to converge always to a valid solution (if it exists). The problem is solved in a higher dimensional space to speed up the convergence.

When a robot has to solve a task in a unconstrained general environment it is required to know both, its position, and the uncertainty for this position estimate. It is usual to express this uncertainty by means of a covariance matrix. In a previous work we presented a method to obtain a close form expression for the uncertainty in the odometry position estimate of a mobile vehicle using a covariance matrix whose form is derived from the kinematic model. We then particularize for a non-holonomic Ackerman driving type autonomous vehicle. Its kinematic model relates the two measures being obtained for internal sensors: the velocity, translated into the instantaneous displacement, and the instantaneous steering angle. However, obtaining a close form expression for the cross-covariance terms between the previous position of the robot and its actual increment of position is not straight forward. Thus, a formulation to obtain a close form expression for these terms is developed. The basic idea of the method herein presented is to fit the covariance matrix for the previous position using a set of equations that, at the same time, simplifies its mathematical expression. A Monte Carlo simulation is performed in order to compare the obtained position uncertainty with the presented method and others used to estimate and propagate the errors, such as the classical Jacobian and the scaled unscented transformation.

In this presentation research activities are discussed, which have been developed at LARM: Laboratory of Robotics and Mechatronics, University of Cassino. After a brief introduction on teaching and research issues developed at LARM, specific attention is addressed to both passive and active cable-based parallel manipulators. In particular, passive cable robots can refer to measuring systems, while active cable-based parallel manipulators are actively actuated. Prototypes and experiences are described as concerning with CATRASYS (Cassino Tracking System), a pose estimation device, and CALOWI (Cassino Low-Cost easy-Operation Wire Robot), a 4-4 cable-suspended manipulator. Further applications are discussed together with those referring to current collaboration between LARM and IRI.

Although the original idea is not new (Neocognitron - Fukushima - 1991), recent neurophysiological findings in the visual cortex have helped to come up with a model capable of interesting invariant capabilities, that competes with some of the best state-of-the-art computer vision systems. The so called standard model was fistly implemented in the MIT (Serre, Wolf and Poggio - 2005) and recently improved (August 2006). The Hierarchical Model progressively processes the visual information through several levels, that need the output of the previous ones to proceed, and are organized hierarchically. It is still unclear, however, which bio-inspired principles are best in computer vision systems; so the objective of the talk will be to present a basic model and to discuss possible changes.

I will expose the ideas behind a couple of works where photometric invariants are used to detect or classify color edges. The first work uses two interesting ideas to classify the physical nature of edges: an automatic noise-adaptive local thresholding; and a taxonomy on color-edge types based on the properties of different color models. The second work proposes a set of quasi-invariants based on derivatives that have good discriminative power and are robust againts noise.

In this talk I will present an interval propagation algorithm for variables in planar and spherical single-loop linkages. Given an interval constraining the allowed values for an input variable, the propagation algorithm finds the set of values for an output variable for which there is at least one solution of the loop equation satisfying the constraint. A recent extension of the propagation algorithm allows the propagation of multiple constraints imposed on different variables at the same time. This provides a powerful tool to find the solution sets of complex multi-loop linkages by iterative methods.

In this talk we will present a numerical method able to isolate all configurations that an arbitrary loop linkage can adopt, within given ranges for its degrees of freedom. The procedure is general, in the sense that it can be applied to single or multiple intermingled loops of arbitrary topology, and complete, in the sense that all possible solutions get accurately bounded, irrespectively of whether the analyzed linkage is rigid or mobile. The problem is tackled by formulating a system of linear, parabolic, and hyperbolic equations, which is here solved by a new strategy exploiting its structure. The method is conceptually simple, geometric in nature, and easy to implement, yet it provides solutions at the desired accuracy in short computation times.

In this talk we present an observability analysis for a number of single camera SLAM systems. The aim is to get a better understanding of the well known intuitive behaviour of these systems, such as the need for triangulation to features from different positions in order to get accurate relative pose estimates. The characterization of the unobservable directions is made using the nullspace basis of the stripped observability matrix. This allow us to identify the type of inputs that should be avoided in order not to incur in unobservable conditions. The analysis is performed modeling the systems in the continous time domain as piecewise linear.

The idea of moving robots or prosthetic devices not by manual control, but by mere thinking (i.e., the brain activity of human subjects) has fascinated researchers for the last 30 years, but it is only now that first experiments have shown the possibility to do so. Such a kind of brain-computer interface (BCI) is a natural way to augment human capabilities by providing a new interaction link with the outside world and is particularly relevant as an aid for physically disabled people. In this talk I will review the field of BCI, with a focus on how brainwaves can be used to directly control robots. Most of the hope for such a possibility comes from invasive approaches that provide detailed single neuron activity; however, it requires surgical implantation of microelectrodes in the brain. For humans, non-invasive systems based on electroencephalogram (EEG) signals are preferable but, until now, have been considered too poor and slow for controlling rapid and complex sequences of movements. Recently we have shown for the first time that online analysis of a few EEG channels, if used in combination with advanced robotics and machine learning techniques, is sufficient for humans to continuously control a mobile robot.

S'analitza la precisió d'un métode per estimar com es belluga una cámera a partir de les imatges que va captant. D'una vista a una altra es produeix una transformació afí del pla de la imatge caracteritzada per 6 parámetres, a partir dels quals es calcula la translació i la rotació de la cámera. Mitjançant simulacions de Monte Carlo es valida l'aplicabilitat d'un procediment estadístic menys costós, la Unscented Transformation, que permet obtenir les mitjanes i covariances del moviment 3D aplicant el mètode només a 13 mostres per vista. Els resultats indiquen que les translacions i la rotació en el pla de la imatge es recuperen millor que la translació segons l'eix òptic, que a la seva vegada supera les dues rotacions fora del pla. Quant a les covariances, només apareix un cert lligam entre els tres graus de llibertat menys precisos.

Grasping and manipulation of objects are fundamental tasks in the use of robots in industrial processes as well as in robotized applications in non-structures environments. The development of versatile end-effectors, like the mechanical hands, has the intention of widening the field of robot applications as well as the efficiency of the robots, but, in contrast, the kinematics and control complexity of these devices and the constraints on the grasping action make it necessary the use of systems that automatically solve the different problems associated with a grasping operation. This talk presents some approaches developed in my thesis to solve two problems involved in grasping and manipulation of objects by means of mechanical hands: the determination of the contact points between the fingers and the object that assure the object immobility, and the determination of the contact forces that each finger has to exert in order to maintain the object equilibrium. In general, these problems can have multiple solutions, some of them being more adequate than others for a given purpose. Then, the objective of the proposed approaches is the determination of the final solution taking into account quality criteria, namely, the robustness of the grasp in front of external perturbations forces and finger positioning errors in the determination of the contact points, and the minimization of the contact forces.

This special seminar will be a brief introduction to the three STREP project proposals in which the IRI robotics group is involved. The three STREP proposlas are:

• URUS: Ubiquitous Robotics in Urban Setting (Alberto Sanfeliu)

• MARIONET: Wire-driven parallel robot for human being servicing (Federico Thomas)

• AIR: Advanced Interactive Robotics (Carme Torras)

Biological agents can in a very flexbile way adapt to their environment and this is achieved by changing the properties of large networks of neurons through synaptic plasticity following learning. Principles of synaptic plasticity have been used in technical fields in conjunction with the design of artificial neural networks (ANNs), but examples where neural networks are actually controlling complex robots are still quite rare. The conflict between the required reliablility in robot control and the more fuzzy and distributed responses of neuronal controllers makes it hard to use ANNs for achieving robust behavior. Here we show that one can design a synaptic learning rule at the single unit level, implement this into individual neurons, which already control simple actions, and finally design a more complex 24-unit neuronal controller to generate robust and adaptable gaits in a planar biped walking robot.

In this talk we will introduce a suite of algorithms to deal with planar linkage configuration spaces. The kernel of this suite is a numerical method able to compute all possible configurations of a planar linkage. The procedure is applicable to rigid linkages (i.e., those that can only adopt a finite number of isolated configurations) and to mobile ones (i.e., those that have internal degrees of freedom). The method is based on the fact that this analysis always reduces to finding the roots of a polynomial system of linear, quadratic, and hyperbolic equations, which is here tackled with a new strategy exploiting its structure. The method is conceptually simple, geometric in nature, and easy to implement, yet it provides solutions of the desired accuracy in short computation times. Experiments are included which show its performance on the double butterfly linkage, for which an accurate an complete discretization of its configuration space is obtained.

This seminar introduces the main topics developed in my thesis integrating knowledge from several fields such as underwater robotics, computer vision, VLSI architectures and reconfigurable devices. Underwater robotics was the motivation of this work, even though computer vision and parallel VLSI architectures played the most important role. Many technologies exist to provide information about a vehicle position, but among them vision based systems represent a good option due to their low cost, high-rate and high-resolution. From the relative observation of the seafloor, a vision system can estimate the vehicle's motion. While moving, the robot can build a map of the seafloor. These visual maps are known as mosaics and can be used for the localisation of the vehicle. The apparent motion of a camera mounted on an underwater vehicle can be estimated by correlating two successive frames of an image sequence. However, searching for correspondences is a time-consuming and error-prone process. Lack of well-defined contours caused by blurring of the elements of the image, as well as non-uniform illumination when using artificial light makes underwater scenes much more difficult to be processed than normal images. Therefore, methods frequently used in standard image processing must be modified and adapted to these particular conditions. We proposed in this thesis a method based on texture characterisation of points to reject outliers from the image correspondence problem. On the other hand, a parallel implementation was used to speed-up parts of the motion estimation algorithm which have a computationally high load. A new VLSI architecture is proposed with the aim of achieving frame-rate performance.

Our research pursuits the autonomous navigation of mobile robots in outdoor environments. In these environments the robot is going to rely on the detection, characterization and later recognition of natural landmarks for orientation and localization purposes. In this context, a color based image segmentation algorithm has been designed to extract the most relevant regions composing a natural landmark. Then, the spatial moments and the topological relations of the extracted regions are used together with their color information for the characterization of the landmark. Our image segmentation strategy combines the color discrimination capabilities of histograms with the spatial capabilities of region growing methods and works in the HSI color space for its robustness to changing illumination conditions. But the color discrimination is not uniform in any of its hue, saturation or intensity domains. Therefore, we have modelled those heterogeneous discrimination properties in order to obtain a uniform color distance measure that provides proper segmentation results. The result is a fast and robust color segmentation algorithm suitable for our robotic navigation system.

In applications where an agent is required to develop a complex task it is hard to code which action must be executed in each situation. It is simpler to make the agent learn to execute the task autonomously, interacting with the environment and learning by the experience. This learning paradigm is denoted as reinforcement learning (RL). The main difficulty of RL is the vast number of different situations the agent should experience in a complex task, making the straight application of this paradigm impracticable in many problems. To make it feasible the agent must be supplied with the capability to generalize from the experienced situations in order to infer which action to execute in those situations never perceived. The same generalization problem is raised in many fields of AI, where some inference is required in problems where the amount of information is computationally intractable. Therefore, many of the generalization techniques applied in RL are the same as those applied in other learning paradigms, mainly in supervised learning. These techniques still have problems when generalizing in complex tasks, constituting an open research area. Our work consists in finding a generalization technique that improves the results obtained so far, and develop a method that combines this technique with RL for an agent to learn complex tasks in real environments. One of the main assumptions in our research is that each variable the agent must perceive to complete a task has different relevancies for the inference depending on the situation, where a situation not only consists of perceptions but also implicates actions. Current generalization techniques assume that a variable has the same relevance for all the situations the agent could face wasting opportunities of generalization. The most relevant perceived variables for a situation, with the coupled actions, are gathered in a rule. In order to create a structure that supports the rules system, where some variables should be considered for some situations and not for others, we developed a function representation that consists in overlapped subspaces of the domain. Each subspace has assigned a value of the represented function (in our case, the cumulative reward q). Given a particular instance (a situation) many overlapped subspaces are involved and only one of them determines the value of the function. This is done after a competition among the involved subspaces where the subspace that wins determines the value of the function. The developed technique progressively builds and embeds rules in the mentioned representation, solving the arising representation inconsistencies while learning the value function. These rules finally contain the relevant information to optimally complete the task. The developed technique so far finds a consistent function representation, deals with the inherent uncertainties of the real environment by using estimation theory, and is robust to nonstationarities of the system as required by the RL.

In this talk I will describe a value iteration algorithm for learning to act in Partially Observable Markov Decision Processes (POMDPs) with continuous state spaces. Mainstream POMDP research focuses on the discrete case and this complicates its application to, e.g., robotic problems that are naturally modeled using continuous state spaces. The main difficulty in defining a (belief-based) POMDP in a continuous state space is that expected values over states must be defined using integrals that, in general, cannot be computed in closed form. We will first see that the optimal finite-horizon value function over the continuous infinite-dimensional POMDP belief space is piecewise linear and convex, and is defined by a finite set of supporting functions that are analogous to the vectors (hyperplanes) defining the value function of a discrete-state POMDP. Second, we will see that, for a fairly general class of POMDP models in which all functions of interest are modeled by Gaussian mixtures, all belief updates and value iteration backups can be carried out analytically and exact. A crucial difference with respect to the vectors of the discrete case is that, in the continuous case, the functions will typically grow in complexity (e.g., in the number of components) in each value iteration. Finally, I will describe the application of Perseus, a randomized point-based value iteration algorithm, in a simple robot planning problem with a continuous domain.

La presentació constarà de dues parts: primer descriuré alguns dels resultats obtinguts a l'IRI en l'aplicació de xarxes neuronals a l'aprenentatge de la cinemàtica i la coordinació visuomotora de robots, i després esbossaré les línies directrius del projecte "Perception, Action and Cognition through Learning of Object-Action Complexes", aprovat dins del 6è Programa Marc de la Comissió Europea i que començarà el proper febrer.

We consider a single hand-held camera performing SLAM at video rate with generic 6DOF motion. The aim is to optimise both the localisation of the sensor and building of the feature map by computing the most appropriate control actions or movements. The actions belong to a discrete set (e.g. go forward, go left, go up, turn right, etc), and are chosen so as to maximise the mutual information gain between posterior states and measurements. Maximising the mutual information helps the camera avoid making ill-conditioned measurements appropriate to bearing-only SLAM. Moreover, orientation changes are determined by maximising the trace of the Fisher Information Matrix. In this way, we allow the camera to continue looking at those landmarks with large uncertainty, but from betterposed directions. Various position and gaze control strategies are first tested in a simulated environment, and then validated in a video-rate implementation. Given that our system is capable of producing motion commands for a real-time 6DOF visual SLAM, it could be used with any type of mobile platform, without the need of other sensors.

We have developed a suite of MatLab scripts for solving the Input/Output problem of planar mechanisms. The mechanism is described in a matrix that is given to the toolbox as input. The toolbox extracts the loop-closure equations and finds all isolated roots of the equation system by polynomial continuation homotopy. Then all the possible configurations of the mechanisms are drawn.

We will describe a system to detect objects on images that is invariant to translations, scaling and rotations of the object, moreover it presents a low computational cost and high detection rates. The system is based on local features to detect contours that are combined with the AdaBoost Algorithm to obtain a robust classifier able to detect the object in different contexts.

This work discusses some aspects of the process of developing of parallel computing machine. The mathematical concept "algebra" is used as a basis of the investigation,. The application of this notion allows the representation of any task of this type as an ordered structure of three components: set of elements (Xilinx's FPGA), set of rules for action (concrete, particular algorithm), and set of constraints. It is shown that such an approach cover in practice the complete set of all valuable varieties of the problem and permits to find the optimal solution for every particular task.

An approach is described aimed at the detection of lines in digital images and evaluation of metric parameters of the objects. It is based on the properties of the Fourier Transform which allow accumulating energy of pixels lying on straight lines of same slope alongside a single straight line in the frequency domain. Using only the spectrum values alongside the latter line permits the detection of straight line segments in the original image and the evaluation of specific parameters. This technique also allows detecting characteristic points like corner points, points of intersection or bifurcation, thus describing the original image as a set of geometric features. The approach could be applied to the problems of document processing, robot vision and like.

We present a probabilistic path planning method which computes collision-free paths for two robotic arms Staübli RX60 that simultaneously grasp an object. The method assumes that the environment is static and that accurate geometric models of their objects are known. It proceeds in two phases: a learning phase and a query phase. In the learning phase, a probabilistic roadmap is constructed and stored as a graph whose nodes correspond to collision-free configurations and whose edges correspond to feasible paths between these configurations. These paths are computed using a simple and fast local planner. In the query phase, any given start and goal configurations of the two robotic arms are connected to two nodes of the roadmap, and the roadmap is then searched for a path joining them. The method is an adaptation of the well-known algorithm by Svetska et al. , which has been appropriately modified so as to take into account the closure constraint imposed by the simultaneous grasp of the object.

Most solutions to the SLAM problem in robotics have utilized Range and Bearing sensors as the provided perception data is easy to incorporate, allowing immediate landmark initialization. This is not the case when using Bearing-Only information because the distance to the perceived landmarks is not directly provided. A whole estimate of a landmark position will only be possible via a set of measurements taken from different points of view.

The vast majority of contributions to this problem perform a parallel task to get this estimate, and hence the landmark initialization is delayed. However, a second family of methods exist that map the bearing information of the landmark without the need of knowing its distance. Pros and cons of both delayed and undelayed families will be discussed, as well as today's most performing algorithms for each one of them in the EKF-SLAM framework.

Pros and cons for vision-based SLAM systems can be resumed as follows: a) delayed methods are more optimal than undelayed ones, in the senses of information loss and consistency; b) only undelayed methods can be used in vehicles that have the vision sensors looking in the motion direction; and c) delayed methods can more easily deal with unstable or short living detected visual features.

The boost in algorithm performance is mainly based on the following hints: a) non-Gaussian PDFs of the unobserved distance to the landmarks are approximated by Gaussian mixtures instead of sets of particles; b) these Gaussian mixtures are best defined as geometric series of Gaussians; c) for the delayed method, several landmarks are tracked at each frame and only the best ones will be initialized; and d) for the undelayed method, the Federated Information Sharing filter updating technique is developed and used. The consequence of these choices is the possibility to simultaneously initialize several landmarks in real time, with both the delayed and undelayed algorithms.

Some experimental results are given to show the pertinence of the proposed methods.

Motion analysis and object tracking has been one of the principal focus of attention over the past two decades within the computer vision community. The interest of this research area lies in its wide range of applicability, extending from autonomous vehicle and robot navigation tasks, to entertainment and virtual reality applications.

Even though impressive results have been obtained in specific problems, object tracking is still an open problem, since available methods are prone to be sensitive to several artifacts and non-stationary environment conditions, such as unpredictable target movements, gradual or abrupt changes of illumination, proximity of similar objects or cluttered backgrounds. Multiple cue integration has been proved to enhance the robustness of the tracking algorithms in front of such disturbances. In recent years, due to the increasing power of the computers, there has been a significant interest in building complex tracking systems which simultaneously consider multiple cues. However, most of these algorithms are based on heuristics and ad-hoc rules formulated for specific applications, making impossible to extrapolate them to new environment conditions.

In this dissertation we propose a general probabilistic framework to integrate as many object features as necessary, permitting them to mutually interact in order to obtain a precise estimation of its state, and thus, a precise estimate of the target position. This framework is utilized to design a tracking algorithm, which is validated on several video sequences involving abrupt position and illumination changes, target camouflaging and non-rigid deformations. Among the utilized features to represent the target, it is important to point out the use of a robust parameterization of the target color in an object dependent colorspace which allows to distinguish the object from the background more clearly than other colorspaces commonly used in the literature.

In the last part of the dissertation, we design an approach for relighting static and moving scenes with unknown geometry. The relighting is performed through an `image-based' methodology, where the rendering under new lighting conditions is achieved by linear combinations of a set of pre-acquired reference images of the scene illuminated by known light patterns. Since the placement and brightness of the light sources composing such light patterns can be controlled, it is natural to ask: what is the optimal way to illuminate the scene to reduce the number of reference images that are needed? We show that the best way to light the scene (i.e., the way that minimizes the number of reference images) is not using a sequence of single, compact light sources as is most commonly done, but rather to use a sequence of lighting patterns as given by an object-dependent lighting basis. It is important to note that when relighting video sequences, consecutive images need to be aligned with respect to a common coordinate frame. However, since each frame is generated by a different light pattern illuminating the scene, abrupt illumination changes between consecutive reference images are produced. Under these circumstances, the tracking framework designed in this dissertation plays a central role. Finally, we present several relighting results on real video sequences of moving objects, moving faces, and scenes containing both. In each case, although a single video clip was captured, we are able to relight again and again, controlling the lighting direction, extent, and color.

The hostile conditions existing beyond the Earth's atmosphere convert space missions into one of the most well-founded applications of robots but also into one of the most challenging ones. A survey on the two big families of space robots is presented here: robotic arms for EVA operations (and their "natural" extension: humanoids), and rovers for planetary exploration (highly popular due the success of the last Mars missions). The talk will conclude with some new robotic concepts that cannot be classified into these categories.

The hostile conditions existing beyond the Earth's atmosphere convert space missions into one of the most well-founded applications of robots but also into one of the most challenging ones. A survey on the two big families of space robots is presented here: robotic arms for EVA operations (and their "natural" extension: humanoids), and rovers for planetary exploration (highly popular due the success of the last Mars missions). The talk will conclude with some new robotic concepts that cannot be classified into these categories.

In some applications, it is useful to retrieve the shape of a surface using a robot manipulator. Usually, this goal is achieved using the encoder data obtained while the robot is following the surface using a hybrid position/force or speed/force control. However, nonlinearities associated with the actuators such as backlash make impossible to get the real contour from the encoder data. To solve this problem, an external position feedback is added. This new feedback loop is based on a PSD (Position Sensing Device) sensor which presents some advantages over a traditional CCD camera such as speed an accuracy. Some experimental results are presented using a 3 dof manipulator following a 2D contour.

A user interface for the control of vision-based robot navigation in previously unknown, indoor or outdoor environments will be described. The user, by means of a visual feedback from the camera(s) of the robot, can select a visual target to be reached by the robot and launch an autonomous navigation process. Manual control can be taken back by the user at any time. The interface is built as a modular platform, capable of accommodating different types of robots and different algorithms for vision and navigation. Thus, in addition to constitute a generic interface for navigation control, the platform can be used as a tool to test different navigation approaches and compare their performances in similar contexts.

An image can be seen as an element of a vector space and hence it can be expressed in as a linear combination of the elements of any non necessarily orthogonal basis of this space. After giving a matrix formulation of this well-known fact, this paper presents a reconstruction method of an image from its moments that sheds new light on this inverse problem. Two main contributions are presented: (a) the results using the standard approach based on the least squares approximation of the result using orthogonal polynomials can also be obtained using matrix pseudoinverses, which implies higher control on the numerical stability of the problem; and (b) it is possible to use basis functions in the reconstruction different from orthogonal polynomials, such as Fourier or Haar basis, allowing to introduce constraints relative to the bandwidth or the spatial resolution on the image to be reconstructed.

Using internal and external sensors to provide position estimates in a two-dimensional space is necessary to solve the localization and navigation problems for a robot or an autonomous vehicle. Usually, a unique source of position information is not enough so researchers try to fuse data from different sensors using several methods as for example Kalman filtering. Those methods need an estimation of the uncertainty in the position estimates obtained from the sensory system. This uncertainty is expressed by a covariance matrix, which is usually obtained from experimental data assuming, by the nature of this matrix, general and unconstrained motion. We propose in this paper a close form expression for the uncertainty in the odometry position estimate of a mobile vehicle using a covariance matrix whose form is derived from the cinematic model. We then particularize for a non-holonomic Ackerman driving type autonomous vehicle. Its cinematic model relates the two measures being obtained for internal sensors: the velocity, translated into the instantaneous displacement, and the instantaneous steering angle. The proposed method is validated experimentally, and compared against Kalman filtering.

This seminar presents a force and torque sensor developed within the project "Multi-arm roboticed cell" ("Estació de treball multibraçs"), funded by the CERTAP, a reference center of the Generalitat de Catalunya, the Catalan government. The sensor has the structure of a Stewart-Gough platform: it is made of two rigid plates, a base and a platform, linked by six legs, each mounting a single-axis load-cell to measure the force on it. We will see how the measures of the six legs can be combined to derive 1) the net wrench (force and torque) acting on the platform, 2) the line of application of the force, and 3) the uncertainties associated with all these magnitudes due to errors in the measured loads. The device is available in our lab for anyone wishing to play with it (just contact any of the authors).
The work presented in this seminar has been done in cooperation with Roger Frigola, Francesc Roure, and Federico Thomas.

Most approaches to camera motion estimation from image sequences require matching the projections of at least 4 non-coplanar points in the scene. The case of points lying on a plane has only recently been addressed, using mainly projective cameras. We here study what can be recovered from two uncalibrated views of a planar object under affine viewing conditions. Using results in projective geometry, we prove that the affine epipolar direction can be recovered provided camera motion is free of cyclorotation. This result is then confirmed analytically, and subjected to experimentation. A setup consisting of a Staubli robot holding a planar object in front of a camera is used to obtain calibrated image streams. The object contour is tracked and its affine deformation between any two frames extracted. The fact that our method and the Gold Standard algorithm produce comparable results shows the potential of our proposal, since with less information (only from a plane) and with a much simpler processing (solving a single second-order equation), we obtain the epipolar direction with similar accuracy.

In this article we propose an algorithm to reduce the effects caused by linearization in the typical EKF approach to SLAM. The technique consists in computing the vehicle prior using an Unscented Transformation. The UT allows a better nonlinear mean and variance estimation than the EKF. There is no need however in using the UT for the entire vehicle-map state, given the linearity in the map part of the model. By applying the UT only to the vehicle states we get more accurate covariance estimates. The a posteriori estimation is made using a fully observable EKF step, thus preserving the same computational complexity as the EKF with sequential innovation. Experiments over a standard SLAM data set show the behavior of the algorithm.

The inverse/direct kinematics of trilaterable serial/parallel manipulators can be stated as a system of distance constraints whose set of solutions can be determined using a sequence of trilaterations, possibly involving points at infinity. It is possible to decide whether a mechanism is trilaterable by relying only on its topology. Based on this fact, we here enumerate all trilaterable serial and in-parallel robots with six degrees of freedom. The relevance of the obtained family of manipulators is established when it is shown to contain the best-known commercial serial robots. As a result of this analysis, we come up with a general method to solve the inverse/direct kinematics of a wide family of manipulators.

It has been shown in the literature that image-based relighting of scenes with unknown geometry can be achieved through linear combinations of a set of pre-acquired reference images. Since the placement and brightness of the light sources can be controlled, it is natural to ask: what is the optimal way to illuminate the scene to reduce the number of reference images that are needed? We show that the best way to light the scene (i.e., the way that minimizes the number of reference images) is not using a sequence of single, compact light sources as is most commonly done, but rather to use a sequence of lighting patterns as given by an object-dependent lighting basis. While this lighting basis, which we call the optimal lighting basis (OLB), depends on camera and scene properties, we show that it can be determined as a simple calibration procedure before acquisition. We demonstrate through experiments on real and synthetic data that the optimal lighting basis significantly reduces the number of reference images that are needed to achieve a desired level of accuracy in the relit images. This reduction in the number of needed images is particularly critical in the problem of relighting in video, as corresponding points on moving objects must be aligned from frame to frame during each cycle of the lighting basis. We show, however, that the efficiencies gained by the optimal lighting basis makes relighting in video possible using only a simple optical flow alignment. We present several relighting results on real video sequences of moving objects, moving faces, and scenes containing both. In each case, although a single video clip was captured, we are able to relight again and again, controlling the lighting direction, extent, and color.

Robotics applications often involve dealing with complex dynamic systems. In these cases coping with control requirements with conventional techniques is hard to achieve and a big effort has to be done in the design and tuning of the control system. An alternative to conventional control techniques is the use of automatic learning systems that could learn control policies automatically, by means of the experience. But the amount of experience required in complex problems is intractable unless some generalization is performed. Many learning techniques have been proposed to deal with this challenge but the applicability of them in a complex control task is still impracticable because their bad learning convergence or insufficient generalization. In this work a new learning technique, that explodes a kind of generalization called categorization, is used in a complex control task. The results obtained show that it is possible to learn, in short time and with good convergence, a control policy that outperforms a classical PID control tuned for the specific task of controlling a manipulator with high inertia and variable load.