Following the Human Thread in Social Navigation

The success of collaboration between humans and robots in shared environmentsrelies on the robot's real-time adaptation to human motion. Specifically, inSocial Navigation, the agent should be close enough to assist but ready to backup to let the human move freely, avoiding collisions. Human trajectories emergeas crucial cues in Social Navigation, but they are partially observable fromthe robot's egocentric view and computationally complex to process. We propose the first Social Dynamics Adaptation model (SDA) based on therobot's state-action history to infer the social dynamics. We propose atwo-stage Reinforcement Learning framework: the first learns to encode thehuman trajectories into social dynamics and learns a motion policy conditionedon this encoded information, the current status, and the previous action. Here,the trajectories are fully visible, i.e., assumed as privileged information. Inthe second stage, the trained policy operates without direct access totrajectories. Instead, the model infers the social dynamics solely from thehistory of previous actions and statuses in real-time. Tested on the novelHabitat 3.0 platform, SDA sets a novel state of the art (SoA) performance infinding and following humans.