Efficient Robot Navigation in Dense Crowds via Causal Reasoning on Spatio-Temporal Graphs

As mobile robot technology advances, safe and efficient navigation in dynamic crowd environments has become a major challenge. Traditional methods often overlook complex interactions between agents and pedestrians, limiting their dynamic obstacle avoidance in dense crowds. Recent methods using deep reinforcement learning and graph convolutional networks have improved robot navigation, but they focus mainly on spatial interactions and neglect temporal dependencies in crowd dynamics. We design a dense crowd navigation method using spatio-temporal graph reasoning, which constructs a agent-pedestrian interaction graph and applies an interaction-aware graph convolutional network to extract spatial features. We propose a Causal Spatio-Temporal Graph Attention Mechanism that models the temporal dependencies of pedestrian interaction dynamics and captures the causal relationships between an agent's state space and its actions. For multi-step spatio-temporal graphs, we extract features using a multi-layer, multi-head attention architecture to effectively encode their temporal structure. This approach yields more accurate predictions of pedestrian motion trends and improves the robot's obstacle-avoidance performance in complex crowd Experimental results show our method outperforms traditional methods in dense crowd scenarios.