TRAINING ROBOTIC MANIPULATORS FOR SAFE OBSTACLE AVOIDANCE: PLANNING, IMITATION, AND REINFORCEMENT

Gao Tianci, Yang Bo, Rao Shengren

This paper proposes a comprehensive method for training a UR5 robotic manipulator to perform safe and efficient movements in environments with randomly placed obstacles. In the first stage, collision-free “expert” trajectories are automatically generated using the OMPL motion planning library. This eliminates the need for manual robot control and ensures high-quality demonstrations. In the second stage, a behavioral cloning approach is used to derive an initial policy capable of reproducing these demonstrations and avoiding collisions. Finally, reinforcement learning with the PPO algorithm is applied to fine-tune and improve the policy based on a reward function that accounts for positioning accuracy, collision penalties, and other factors. This strategy unites formal trajectory planning with the adaptability inherent to reinforcement learning methods. Experimental results in the simulator show that the proposed three-stage framework — planner–imitation learning–reinforcement learning — achieves higher goal accuracy and a lower collision rate compared to classical approaches trained from scratch.

Keywords:Motion Planning (OMPL); Imitation Learning; Reinforcement Learning; Safe Obstacle Avoidance; Collision-free Trajectory; Industrial Robotics.