Research on Path Planning of Mobile Robots Based on Dyna-RQ

Abstract

The mobile robot path planning problem is one of the main contents of reinforcement learning research. In traditional reinforcement learning, the agent obtains the cumulative reward value in the process of interacting with the environment and finally converges to the optimal strategy. The Dyna learning framework in reinforcement learning obtains an estimation model in the real environment. The virtual samples generated by the estimation model are updated together with the empirical samples obtained in the real environment to update the value function or strategy function to improve the convergence efficiency. At present, when reinforcement learning is used for path planning tasks, continuous motion cannot be solved in a large-scale continuous environment, and the convergence is poor. In this paper, we use RBFNN to approximate the Q-value table in the Dyna-Q algorithm to solve the drawbacks in traditional algorithms. The experimental results show that the convergence speed of the improved Dyna-RQ algorithm is significantly faster, which improves the efficiency of mobile robot path planning.