PI2-BDMPs in Combination With Contact Force Model: A Robotic Polishing Skill Learning and Generalization Approach

Robot skill learning and generalization in robotic polishing tasks play a pivotal role in efficient task planning, particularly in tasks where visual scanning is not convenient. Compared to other manipulations, skill learning of robot polishing trajectories and skill generalization between similar trajectories require high accuracy, and skill learning under a small number of basis function settings is also worth studying. To address these challenges and consider both the learning accuracy and the number of basis function of skill model, this study introduces a novel method known as B-spline dynamic movement primitives (BDMPs) in the context of robotic polishing tasks. BDMPs demonstrate superior trajectory learning accuracy even with a small number of basis functions compared to dynamic movement primitives. In addition, to facilitate trajectory generalization in new scenarios, a trajectory skill generalization framework based on policy improvement with path integrals (PI2) is proposed, resulting in an integrated approach named PI2-BDMPs. The PI2-BDMPs approach is designed to enable the generalization of polished trajectory skills to meet new task requirements, even when provided with limited demonstration points and a small number of basis functions. In order to enhance the generalization performance of the skill in the contact force setting task, a PI2 approach incorporating a contact force model is proposed. To validate the effectiveness of our proposed method, the real-world robot polishing experiments are conducted with an adaptive variable impedance controller. The results verify that BDMPs offer superior trajectory learning accuracy with few basis functions, and PI2-BDMPs facilitate the generalization of polished trajectory skills to new scenarios, showcasing the adaptability of our approach.