Speed profile optimization through directed explorative learning

In this paper we propose a new skill learning framework based on fusing prior knowledge with programming by demonstration and explorative learning methodologies. Prior knowledge as well as all partially known models guide the search process within the proposed adaptation method. The proposed methodology is based on algorithms originating in iterative learning control and reinforcement learning. The developed approach was experimentally verified on the problem of speed profile optimization for a challenging task of transferring vessels filled with liquid without spilling. In order to explicitly encode the speed profiles and to allow their transfer between tasks, a modified form of dynamic movement primitives has been developed.