A SAC-Bi-RRT Two-Layer Real-Time Motion Planning Approach for Robot Assembly Tasks in Unstructured Environments

摘要

Due to the uncertainty and complexity of the assembly process, the trajectory planning of a robot needs to consider the real-time obstacle avoidance problem when it completes the assembly in the unstructured workspace. To realize the safe assembly of assembly robots in dynamic and complex environments, a dynamic obstacle avoidance trajectory planning method for robots combining traditional planning algorithms and deep reinforcement learning algorithms is proposed to improve the robot’s agent and obstacle avoidance ability in dynamic and complex environments. The Bidirectional Rapidly-exploring Random Tree (Bi-RRT) method is utilized as a global planner to plan the global optimal path quickly; considering the real-time nature of the assembly process, the Soft Actor-Critic (SAC) is used as a local obstacle avoider to avoid obstacles more accurately and to find the nearest node generated by the Bi-RRT during the planning process, which is regarded as the goal during the local obstacle avoidance to reduce the model’s complexity. By training and testing in the simulation engine and comparing with SAC, DDPG and DQN algorithms, the method can avoid obstacles in dynamic and complex environments more efficiently, which verifies that the proposed hybrid method can accomplish the high-precision planning task with a high success rate.