Multi-scene trajectory planning of robot arm based on DQN algorithm

Abstract To overcome the challenges of low accuracy and poor adaptability to high-dimensional spaces in deep reinforcement learning algorithms, presents an improved algorithm that integrates the APF method with the Q-network. The aim is to improve the obstacle avoidance efficiency of robotic arms in complex environments. By introducing a novel reward , avoid obstacles is greatly enhanced. Furthermore, the incorporation of a competitive network, mitigates the overestimation issue of the value function. This results in faster convergence, reduced training time, and guarantees the optimality of the trajectory. Experimental shows the improved algorithm significantly change the navigation capabilities of robotic arms in various environments, offering a new approach for trajectory planning in complex settings.