Reinforcement learning policies for robots navigating around people often assume that pedestrians are non-reactive, meaning they do not change their behavior in response to the robot. This assumption prioritizes safety but overlooks the fact that real pedestrians differ: some notice and react to the robot, while others do not. According to a new arXiv paper, policies trained without conditioning on this local reactivity may therefore be unnecessarily cautious, limiting the robot's efficiency and naturalness.
The authors introduce a method called reactivity-conditioned residual adaptation, which adjusts the robot's actions based on whether nearby pedestrians are likely to respond. Instead of a one-size-fits-all policy, the robot learns when to adapt its behavior to the observed reactivity of people around it. This is framed as a way to handle distribution shift between training and real-world human behavior.
The abstract does not provide experimental details, so the reported benefits are not yet verified. But the core idea addresses a known gap in human-robot interaction: treating all pedestrians as non-reactive can lead to overly defensive robot behavior. By conditioning on reactivity, the approach aims to make robot navigation both safer and more socially aware.