Active Collision Avoidance for Robotic Arm Based on Artificial Potential Field and Deep Reinforcement Learning

Abstract

To address the local minimum issue commonly encountered in active collision avoidance using artificial potential field (APF), this paper presents a novel algorithm that integrates APF with deep reinforcement learning (DRL) for robotic arms. Firstly, to improve the training efficiency of DRL for the collision avoidance problem, Hindsight Experience Replay (HER) was enhanced by adjusting the positions of obstacles, resulting in Hindsight Experience Replay for Collision Avoidance (HER-CA). Subsequently, A robotic arm collision avoidance action network model was trained based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) and HER-CA methods. Further, a full-body collision avoidance potential field model of the robotic arm was established based on the artificial potential field. Lastly, the trained action network model was used to guide APF in real-time collision avoidance planning. Comparative experiments between HER and HER-CA were conducted. The model trained with HER-CA improves the average success rate of the collision avoidance task by about 10% compared to the model trained with HER. And a collision avoidance simulation was conducted on the rock drilling robotic arm, confirming the effectiveness of the guided APF method.