Research Assistant

Project

In our daily life we are usually grasping objects very often. Most of the times we are not even thinking about questions like "How can I grasp an object?" or "What can I do with a cup?". But if we transfer the task of grasping an object to a robot, its complexity becomes clear.

My work within the EU-Project "Emergence of Cognitive Grasping through Introspection, Emulation and Surprise" (GRASP) deals with the acquisition of manipulation-relevant object knowledge.

The robot observes a human demonstrator, who is manipulating an object. But how does the robot know what to observe in the whole scene? First, the robot has to identify the object candidates, which can be manipulated. The human selects one object for the manipulation by grasping it. Now, the robot knows the interesting object. The 6DoF-trajectory of this object can be observed with a vision-based system. It is important to notice our object-centric point of view: It enables the estimation of the spatio-temporal object properties for manipulation tasks (Read more).

After the extraction of the manipulation-relevant object properties, the question of knowledge representation arises. The representation influences the future abilities of the robot: First, the robot should be able to reuse its knowledge. Second, the knowledge has to be extendable, since the robot can be confronted with new situations. A further important aspect is the question who or what influences the manipulation-relevant object properties. The properties are, of course, not only set by the object and its state, but also by the environment. Therefore, the functionality of an object is also influenced by its environment. (Read more).

Links

Estimation of Spatio-Temporal Object Properties for Manipulation Tasks from Observation of Humans

Representation of Manipulation-Relevant Object Properties and Actions for Surprise-Driven Exploration

Machine Vision and Perception Group

EU FP7 GRASP project

Teaching

Exercise Robotik

Publications