Conceptual Similarity as a Key to High-Level Robot Programming by Demonstration

Abstract:
In the future robots must be able to learn behaviors from human demonstration to solve everyday tasks. Researchers in the field of programming by demonstration (PbD) have gained much insight at the level of motor primitives, teaching a robot to perform generic low-level motions. On the other hand, learning behaviors on a higher abstraction level (high-level PbD) is still a challenge. We focus on learning task goals from demonstration and then apply a symbolic planner to find an action sequence from a new situation to the goal. The difficulty is to derive a symbolic description of task goals from subsymbolic perception. As a solution we present a new high-level learning method based on conceptual spaces. Area of conference topics / keywords: cognitive robotics, learning, programming by demonstration