TS-RIL: A two-stage robot imitation learning framework with motion trajectory learning and obstacle avoidance in real-world operating scenarios

Robot imitation learning is an important research and application direction in the field of robot-based intelligent manufacturing. This technology aims to enhance the autonomous and robustness of robot manipulation in unstructured environments. In this paper, we propose a two-stage robot imitation learning framework with motion trajectory learning and obstacle avoidance, named TS-RIL. Specifically, the proposed TS-RIL consists of a motion trajectory learning algorithm based on the region-constrained dynamic motion primitives (RC-DMPs) and an obstacle avoidance algorithm based on the smooth rapidly-exploring random tree star (S-RRT*), and its basic idea is to divide the robot imitation learning into two stages: motion trajectory learning stage and obstacle avoidance stage. In the motion trajectory learning stage, we design a coupling term based on artificial potential field to extend DMPs to RC-DMPs, and generate the learning trajectories in the constrained regions by real-time updating the learning parameters such as the weights of the forcing term, scaling factor and convergence factor. Then in the obstacle avoidance stage, we propose the S-RRT* algorithm to search for a smooth motion trajectory in the local obstacle region, and recombine it with the trajectory generated by RC-DMPs to generalize a new collision-free motion trajectory. Finally, we further develop the continuous RC-DMPs system architecture, which enables robot trajectory learning, obstacle avoidance and motion execution can be performed sequentially and alternately. To evaluate the overall performance of the proposed TS-RIL, we develop an algorithm verification platform based on the Robot Operating System (ROS) and conduct a series of typical prototype experiments in complex real-world scenarios. The experimental results demonstrate that our TS-RIL can significantly improve the effectiveness and robustness of robot motion trajectory learning and generalization, and outperforms the existing robot motion trajectory learning methods in terms of efficiency, path length and success rate.