A Path Planning Method Based on Deep Reinforcement Learning with Improved Prioritized Experience Replay for Human-Robot Collaboration

Owing to its ability to integrate human flexibility with robotic automation, human-robot collaboration possesses tremendous potential in intelligent manufacturing. A quintessential characteristic of this collaboration is the necessity for robotic arms to cooperate with humans in a dynamically changing environment, wherein humans could be considered as dynamic obstacles. One of the significant challenges in human-robot collaboration is the development of obstacle avoidance strategies for robotic path planning within dynamically changing environments. The inability of traditional two-dimensional path planning methods to handle high-dimensional spaces, therefore, many researchers have turned their attention to deep reinforcement learning, and many deep reinforcement learning methods have been applied to robotic arm path planning. However, most deep reinforcement learning models for robotic arm path planning require a significant amount of training time to achieve convergence. In this study, we introduce an algorithm that synergizes Soft Actor-Critic (SAC) with an improved version of Prioritized Experience Replay (PER)—SAC-iPER. We prioritizes experiences based on task-rewards, employing metrics such as time consumption and collision occurrences, in addition to task completion, to rank experiences. This reward-based ordering significantly boosts the learning process in both speed and quality. The results of this study significantly enhanced the training efficiency of deep reinforcement learning models for robotic arm path planning within human-robot collaboration, paving the way for the development of more efficient human-robot collaborative systems.