Following the Human Thread in Social Navigation

The paper proposes a Social Dynamics Adaptation model (SDA) for Social Navigation, which involves a robot's ability to navigate human-centric environments while maintaining a safe distance and adhering to social norms. The key challenge is to process human trajectories, which are partially observable from the robot's perspective and complex to compute. The proposed SDA model uses a two-stage Reinforcement Learning framework: the first stage involves learning to encode human trajectories and the second stage infers social dynamics from the robot's state-action history. This approach has been tested on the Habitat 3.0 platform, achieving state-of-the-art performance in finding and following humans.