Collision-free Path Planning For Welding Manipulator Via Deep Reinforcement Learning

In the narrow industrial welding scene, it is difficult for the 6DOF manipulator to realize intelligent obstacle avoidance planning. This paper proposes an adaptive reinforcement learning on the path planning method of welding manipulator to find a collision-free path in the limited scene. The sub-actor network is designed to conduct guided search on the main actor-network to achieve effective obstacle avoidance. The overestimation of Q value is alleviated by embedding the return distribution function into maximum entropy to replace the shear-double Q learning of SAC. We evaluate our approach on a group path planning experiment, the experiments demonstrate that our method increases learning efficiency and obtains safer policies.