Dr. Babette Dellen
Contact
Projects
Publications
IRI
Research Interests
My main research goal is to understand how artificial agents can make sense of visual sensory data and interact successfully with the environment, e.g., during the course of an object-manipulation task. Humans have no problem to understand the consequences of their own actions and those of others, allowing them to collect world knowledge through observation, including self-observation. Artificial agents however are still far away from being able to tackle such complex tasks in an autonomous way. One difficulty lies in the fact that the visual world can take vastly different appearances in an image. Objects may look different dependent on their pose, lighting conditions and distance from the observer. Actions can be executed in many different ways, hence movement trajectories can show large variations. This variability of visual data and also its visual content constitutes a major challenge to robotics research. The main focus of my research is the development robust visual representations which can bridge this gap and provide a handle for the robot to interact with its environment. This requires the combination of algorithms operating at different levels of a processing hierarchy, consisting of
1. Data acquisition and data fusion, e.g. using color cameras and depth sensors (stereo, time-of-flight, structured light)
2. Extraction of low-level visual descriptors, e.g. optic flow
3. Grouping and tracking
4. Graph based representations for recognition and learning
The generation of robust visual representations that can be used in robotic object manipulation tasks
requires the incorporation of 3D information, which is needed to describe the spatial relations of objects in a 3D space, but also for executing robotic actions, e.g. grasping movements. The development of algorithms for depth computation and the integration of information from depth sensors play thus an important role in my research.
