Research on path planning optimization of mobile robots empowered by deep reinforcement learning

Aiming at the problems of slow convergence speed and poor obstacle avoidance strategy of deep deterministic policy gradient learning (DDPG) path planning algorithm in dynamic environments, an improved optimization scheme of deep deterministic policy gradient learning based on multi-step expected temporal difference method is proposed. Firstly, the n-step expected temporal difference method is used to solve the optimal strategy to solve the problem of poor training decision function. Secondly, when solving the optimal strategy, an expected variable is introduced and a baseline is set to reduce variance operation to enhance the convergence of the algorithm. Then, a continuous reward function is designed to improve the utilization of rewards by the algorithm, and the storage method of priority experience replay is used to optimize the convergence speed. Finally, a ROS simulation environment is built to verify the above method. Experiments show that compared with traditional algorithms, the proposed algorithm improves the convergence speed by 15%; increases the successful number of cumulative arrivals at the target point by 21.6%; and reduces the probability of collision with dynamic obstacles by 11%.