End-to-end Autonomous Navigation Approach for Multiple Mobile Robots Integrating Attention Mechanism and Velocity Obstacle Method

For the problems of slow convergence speed, poor stability, and weak interpretability in existing end-to-end autonomous navigation algorithms for multiple mobile robots based on deep reinforcement learning (DRL), an end-to-end autonomous navigation strategy for multiple mobile robots utilizing attention-velocity obstacle-soft actor critic algorithm (Attention-VO-SAC) is proposed. Firstly, an attention module is integrated into the network structure of the DRL algorithm, which dynamically adjust state feature weights to focus on key task-related features. Secondly, a new reward function is designed to guide robots to avoid unsafe actions with integration of velocity obstacle method, through constraining the velocity direction of the mobile robot. Thirdly, the influence of state features at different decision-making moments on the algorithm decisions is analyzed through the visualization of attention weight coefficients, enhancing the transparency and interpretability of the proposed algorithm. Finally, comparative simulation and experiments are performed to validate the performances of the proposed algorithm. Simulation results indicate that the proposed Attention-VO-SAC algorithm can achieve fastest convergence with over 80% success rate in all scenarios, demonstrating better stability. Experimental results in real scenes show that the proposed algorithm can complete autonomous navigation tasks for multiple mobile robots.