Description

Robotics has been recently challenged by new tasks, such as the manipulation of deformable objects like cloth, arising from the use of robots outside standard industrial scenarios. These manipulation skills have the goal to assist humans in their daily activities, with the ultimate objective of improving their living standards, especially in the case of elderly or physically impaired people.

In such a context, programming the behavior of a robot may need to be done by non-robotics experts, and take into consideration additional parameters besides the plain trajectories to be followed, such as the stiffness, or impedance, of the robotic manipulator. In this sense, classic control policies can be combined with machine learning algorithms, achieving a data-driven control pipeline. The desired robot's behavior can be inductively inferred from a suitable dataset, without requiring low-level programming of the robot's motions. Specifically, the task skills can be learned: in a prescriptive way, by directly recording their execution by a human expert (the so-called learning from demonstration); in a model-based fashion, by constructing a surrogate model of the robot's surroundings, and letting the robot decide how to act based on it (in a model predictive control fashion).

This thesis proposes advances in the field of data-driven control for robotic manipulation, tackled from both the learning from demonstration, and the model-based perspective. Most of the learning modules involved in the thesis rely on kernel methods. While being computationally costly, kernel methods enjoy strong optimality guarantees, and allow for fast, provably accurate approximations to be deployed.

In this thesis, we first study how to learn time-dependent stiffness profiles from human demonstrations, and design a novel, active learning algorithm to achieve this goal. Besides this contribution, we investigate how the human demonstrations of a given task can be processed efficiently, by means of an approximated, kernel-based motion primitive. On the other hand, considering the model-based scenario, we first propose the usage of efficient predictive control techniques for cloth manipulation. Moreover, we develop an approximated, kernel-based regression algorithm for learning and controlling nonlinear dynamics, within the context of the so-called Koopman operator regression. Besides these algorithmic contributions, we further put a focus on the theoretical properties of the kernel approximations considered, bridging techniques from the statistical learning and the control theory fields. In particular, we prove new rates of convergence both for our efficient motion primitive representation, and for the dynamics model retrieved in Koopman operator regression.

The work is under the scope of the following projects:

• CLOTHILDE: Cloth manipulation learning from demonstration (web)