Robot Crowd Navigation Based on Spatio-Temporal Interaction Graphs and Danger Zones.

One of the main challenges in mobile robotics is achieving safe and efficient navigation in crowded environments. Previous work in robot crowd navigation has primarily focused on all pedestrians and assumed that the dynamics of all agents are known in simulation scenarios. However, in partially observable real-world crowd environments, the performance of existing methods deteriorates rapidly and may even result in the Frozen Robot Problem. To address these challenges, we propose an attention mechanism based on spatio-temporal interaction graphs to capture cooperative interactions between the robot and other agents for navigation decision-making in partially observable environments. To encourage the robot to stay away from potential freeze areas, we construct a danger zone based on pedestrian motion characteristics, which defines the constrained motion space for the robot. We train our network using model-free deep reinforcement learning without any expert supervision. Experimental results demonstrate that our model outperforms state-of-the-art methods in challenging scenarios of partially observable crowd navigation.