PUB - Publikationen an der Universität Bielefeld

Many motor skills have an intrinsic, low-dimensional parameterization, e.g. reaching through a grid to different targets. Repeated policy search for new parameterizations of such a skill is inefficient, because the structure of the skill variability is not exploited. This issue has been previously addressed by learning mappings from task parameters to policy parameters. In this work, we introduce a bootstrapping technique that establishes such parameterized skills incrementally. The approach combines iterative learning with state-of-the-art black-box policy optimization. We investigate the benefits of incrementally learning parameterized skills for efficient policy retrieval and show that the number of required rollouts can be significantly reduced when optimizing policies for novel tasks. The approach is demonstrated for several parameterized motor tasks including upper-body reaching motion generation for the humanoid robot COMAN.