SCHST: Social Cross-Historical-Spatio Transformer for Crowd Navigation

The primary challenge in robot crowd navigation lies in enabling robots to anticipate and adapt to inherently unpredictable and uncooperative nature of pedestrian behaviors. However, previous methods predominantly emphasize modeling social intentions in isolation, thereby neglecting the crucial factor of pedestrians’ forthcoming behavioral changes in response to the robot’s decision. To address this limitation, we propose a Social Cross-Historical-Spatio Transformer (SCHST) for robot crowd navigation. This model maps historical trajectories of the robot and pedestrians into the current robot space through a cross-modal attention, further capturing the potential inherent inter-dependency correlation between pedestrians’ latent social intentions and robot making-decision. Then, incorporating the SCHST model to deep reinforcement learning framework raises farsightedness of the robot motion. Through extensive simulation, we demonstrate that our method outperforms baselines across key metrics, including success rate and navigation efficiency. Finally, real-world deployment validates our effectiveness and practical applicability in real scenarios.