PUB - Publikationen an der Universität Bielefeld

Modern robotic applications create high demands on adaptation of actions with respect to variance in a given task. Reinforcement learning is able to optimize for these changing conditions, but relearning from scratch is hardly feasible due to the high number of required rollouts. We propose a parameterized skill that generalizes to new actions for changing task parameters, which is encoded as a meta-learner that provides parameters for task-specific dynamic motion primitives. Our work shows that utilizing parameterized skills for initialization of the optimization process leads to a more effective incremental task learning. In addition, we introduce a hybrid optimization method that combines a fast coarse optimization on a manifold of policy parameters with a fine grained parameter search in the unrestricted space of actions. The proposed algorithm reduces the number of required rollouts for adaptation to new task conditions. Application in illustrative toy scenarios, for a 10-DOF planar arm, and a humanoid robot point reaching task validate the approach.