Improved Residual Reinforcement Learning for Dynamic Obstacle Avoidance of Robotic Arm

This paper proposes an improved residual deep reinforcement learning method for robot arm dynamic obstacle avoidance and position servo. The proposed method first simplifies the state space by constructing key points and pre-trains a model capable of completing obstacle avoidance tasks using the simplified state. Then, when training with real state information, a guiding network is used to help accumulate good samples, which improves the training efficiency. To overcome the convergence difficulty of residual DQN in robot arm obstacle avoidance, this paper incorporates the action of the feedback controller into the action space and uses incremental reward values to evaluate the actions. Simulation results demonstrate that the proposed method can effectively achieve robot arm dynamic obstacle avoidance and position servo.