A Policy Searched-Based Optimization Algorithm for Obstacle Avoidance in Robot Manipulators

Most current learning from demonstration algorithms are only applied in static work environments. However, using the learned skills directly in a changing environment may lead to unpredictable situations, such as collisions. Therefore, robots not only need to learn skill models from demonstrations but also need to generalize skills in response to the changing environment. This article proposes an relative entropy policy search-based obstacle avoidance algorithm, which naturally incorporates skill demonstration into an optimization problem that minimizes the difference between the optimal trajectory distribution and the demonstration distribution while avoiding obstacles. By imposing upper bound constraints on policy deviation in the optimization problem, the problems of policy jump and policy nonconvergence during policy update are solved. Since the optimization process is done offline, no gradient information is required, which allows the algorithm to avoid the complexities of online gradient computation. Finally, according to the optimal trajectory distribution, the kernelized movement primitive algorithm is used to encode skills, and combined with other task requirements to further achieve skill generalization. The simulation on the LASA dataset and robot obstacle avoidance experiment validate the effectiveness of the proposed method.