Researchers use graph representations to teach robots socially compliant navigation

Autonomous mobile robots need to navigate in a socially compliant manner to minimize disturbance to pedestrians. Current methods face limitations: Reinforcement Learning requires hand-crafted reward functions that oversimplify social behavior, while Imitation Learning directly trains from real-world data but struggles to model social interactions and suffers from error accumulation. To address these issues, Berke Kartal and colleagues at Boğaziçi University introduce an imitation learning framework that leverages spatiotemporal dynamics for socially compliant navigation. Their approach uses a graph-based auxiliary network to encode crowd states by attending to individual pedestrians, capturing both spatial and temporal characteristics of real-world interactions.

The framework also includes a navigation module that mitigates error accumulation by incorporating encoded state predictions and employing a trajectory-level learning objective rather than step-by-step imitation. On simulation and a real-world dataset, the method outperforms established data-driven baselines across various social metrics like pedestrian disturbance and path efficiency. This research bridges the gap between capturing complex social interactions and practical deployment, bringing us closer to robots that can seamlessly navigate busy sidewalks, hospitals, or shopping malls without feeling intrusive.