H2INT: Human-Human & Human-Robot Interaction Transformer for Dense Crowd Navigation

Navigation in crowded environments often falters because algorithms assume uniform pedestrian reciprocity. H2INT addresses this interaction uncertainty through a reinforcement learning framework powered by a two-stage gated Transformer and recurrent policy. Instead of relying on explicit responsiveness labels, the agent infers pedestrian cooperation directly from relative positions. Validated in simulation curricula with decreasing responsiveness and deployed on physical hardware, the approach sustains safe trajectories across varying crowd densities without retraining.