Robot Motion Skills Acquisition Method Based on GU-ProMPs and Reinforcement Learning

This Aiming at robot motor skill acquisition problem, this paper will propose a GP-UKF probabilistic movement primitives (GU-ProMPs) learning frarnework, which combines with the learning from demonstration (LfD) and the policy improvements with path integrals (PI2). Specifically, when the parameterized model of ProMPs cannot be expressed linearly, we will replace the classical ProMPs linear representation with GU-ProMPs to acquire a nonlinear policy, as a result, the representation of imitation learning and the robustness of system are enhanced. Moreover, when the strategy combined with PI2, the tasks of adding additional constraints to the index set can be acquired automatically and completed with high quality. Based on the NAO and UR5 robot, the experimental results of classical proves the effectiveness and feasibility of the method.