Model Predictive Control-based dynamic movement primitives for trajectory learning and obstacle avoidance

摘要

Dynamic Movement Primitives (DMP) is a well-established framework in Learning from Demonstration, widely used for acquiring and reproducing motion patterns. To improve DMP’s adaptability and generalization in complex environment, this paper introduces a novel DMP obstacle avoidance approach by integrating Model Predictive Control (MPC). The proposed method formalizes DMP within an MPC framework and employs the Proximal Averaged Newton for Optimal Control solver to incorporate obstacle avoidance conditions into the cost function, adapting to various environments and improving obstacle avoidance efficiency. Evaluations were conducted in scenarios involving static and dynamic obstacles across two- and three-dimensional spaces, as well as multi-obstacle and dynamic environments. It adapts to the scaling characteristics of DMP, demonstrating robustness in complex scenarios. Additionally, the inclusion of a safety threshold enhances the method’s robustness, ensuring safe obstacle avoidance in uncertain environments. Quantitative comparisons with Artificial Potential Field and Steering Angle approaches highlight superior performance in trajectory efficiency and obstacle avoidance success. Finally, validation through simulations and real-world experiments, including a pick-and-place task on UR5 robot, demonstrate the approach’s practical applicability in robotic operations. • Integrated MPC with PANOC solver into DMP for adaptive trajectory learning. • Enabled robust obstacle avoidance in dynamic, complex environments. • Outperformed APF and Steering Angle in obstacle avoidance via quantitative tests.