General-purpose robots require efficient algorithms for skill acquisition to allow them to effectively solve novel tasks. A promising approach to transfer new, complex skills to robots is by leveraging human knowledge and expertise through learning from demonstrations, while incorporating additional information such as object poses or geometries.

Despite the popularity that learning from demonstration (LfD) has gained over the years, skill generalization has traditionally been a challenging problem. We therefore also investigate topics which allow interactive and incremental imitation learning on robots to improve generalization and enabling online skill modulation through human feedback. This contains via-points to adapt skills locally by modulating demonstrated trajectories, and the usage of so-called task-parameterized models that encode movements with respect to different coordinate systems.

This course gives you the chance to implement novel algorithms making use of prerecorded data. Teams achieving especially promising results will be given the opportunity to implement their algorithms on a real robot. The course will be conducted in close collaboration with the Interactive Skill Learning group at DLR.

TUMonline entry: campus.tum.de/tumonline/pl/ui/$ctx/wbLv.wbShowLVDetail