Reinforcement learning-driven dynamic obstacle avoidance for mobile robot trajectory tracking

In this work, we propose a trajectory tracking method based on optimized Q-Learning (QL), which has realtime obstacle avoidance capability, for controlling wheeled mobile robots in dynamic local environments. Based on the observation data and the state of the robot, the designed reinforcement learning (RL) method can determine the obstacle avoidance action during trajectory tracking while simultaneously utilizing controllers to maintain action precision. Through a simple observation space data processing method (OSDPM), the inputting data from the equipped raw lidar is transformed into a dimensionality reduction index vector containing the surrounding environment information of the mobile robot, which can guide QL to quickly correspond the current observation state of the robot to the table state of the QL. To improve the iteration and decision efficiency of the RL method, we optimize the Q -Table structure based on the type of data used. Finally, the simulation results verify the effectiveness of the OSDPM and the obstacle avoidance ability of RL method in unknown local environment.