Robot Navigation Through Crowded Dynamic Scenes with Transformer - DRL

This paper presents a robot navigation method that integrates the Transformer model with Deep Reinforcement Learning (DRL) for autonomous navigation in crowded and dynamic environments. By incorporating pedestrian kinematics features, 2D lidar features, and sub-goal points into an end-to-end navigation network based on the PPO algorithm, the method leverages the multi-head attention mechanism of the Transformer to efficiently process environmental information, generating effective motion commands. Experimental results show that this approach achieves higher success rates than existing methods in most test scenarios and also performs well in terms of average speed, demonstrating its strong generalization capability in complex and unknown environments.