Attention-Based Reinforcement Learning with Center Reward Classification for Redundant Manipulators Motion Optimization

Reliable and efficient obstacle-avoidance motion planning for redundant manipulators remains challenging, especially in environments with irregular obstacles and high-dimensional constraints. Although deep reinforcement learning (DRL) offers promising solutions, existing methods still suffer from slow convergence and suboptimal trajectory quality. This paper accounts for practical path-length and time constraints and proposes an improved DRL-based approach that exhibits significantly faster convergence. Firstly, an attention mechanism is incorporated into the deep reinforcement learning framework to improve the model’s ability to capture spatial features in complex environments. Secondly, a novel experience replay mechanism is proposed to enhance the effective utilization of offline trajectories, thereby substantially accelerating the training process. Thirdly, a parallel experience replay buffer with temporal and path-length constraints is designed, enabling further policy refinement once the robotic manipulator consistently reaches the target position. Experimental results demonstrate that our method achieves significantly shorter paths and lower completion times in complex environments characterized by irregular obstacles.