Layered Programming by Demonstration and Planning for Autonomous Robot Manipulation

We propose a layered system for autonomous planning of complex service robot environment manipulation challenges. Motion planning, logic-based planning and probabilistic mission planning are integrated into a single system and planning models are generated using Programming by [human] Demonstration (PbD). The strength of planning models arises from the flexibility they give the robot in dealing with changing scenes and highly varying sequences of events. This comes at the cost of complex planning model representations and generation, however. Manually engineering very general descriptions covering a large sets of challenges is infeasible as is learning them exclusively by robot self-exploration. Thus, we present PbD for planning models together with generation of parameters from analysis of geometric scene properties to tackle that difficulty. Experimental results show the applicability of these techniques on natural learning and autonomous execution of complex robot manipulation challenges.