Physics-informed spatio-temporal graph reinforcement learning for safe crowd navigation
Standard distance-based reinforcement learning (RL) policies often fail in dense, mixed-speed crowds because they evaluate pedestrian threats based primarily on spatial proximity, ignoring actual collision timing. We introduce PSTGNav, a physics-informed navigation framework that embeds kinematic tracking into the RL perception and control loop. Instead of learning spatial correlations from scratch, our Physical Spatio-Temporal Graph Encoder computes spatial adjacency using a Time-to-Collision (TTC) metric. This biases the network to prioritize fast-approaching targets over closer but safer pedestrians. A kinematic attention bottleneck then condenses these varying neighbor states into a fixedsize environment vector. To prevent reactive freezing behaviors, we train the policy using Horizon-Aware Proximal Policy Optimization (HA-PPO), which applies a short-term constant-velocity projection penalty to penalize unsafe steering choices before a physical crash occurs. Evaluations across both standard uniform-speed benchmarks and heterogeneous speed stress-tests confirm that PSTG-Nav lowers safety buffer intrusions and mitigates deadlocks. Ablation studies demonstrate that achieving this robustness requires the synergy of explicit physical priors and an optimized spatio-temporal neural architecture; this synergy enables our method to enhance kinematic safety without compromising navigation efficiency.