PERSCO: Perceived safe control of mobile robots in human groups with active learning

Collision-free motion does not guarantee that humans perceive a robot as safe; even trajectories that avoid collisions can feel threatening. Conversely, overly cautious behavior, such as reducing speed excessively, can compromise task performance. We present PERSCO, a framework for perceived-safe control of mobile robots that adapts robot behavior in real time to nearby people. We formulate the adaptation as a multi-objective optimization over two key safety parameters: robot speed and distance to humans, and propose an active learning approach that continuously updates these parameters using less obtrusive human feedback (e.g., simple visual signals). Simulation experiments show that PERSCO estimates perceived safety parameters with high accuracy and achieves lower rates of unsafe situations compared to classical active learning and black-box optimization methods. A human-subjects study ( $$N=54$$ N = 54 ) with an autonomous mobile robot demonstrates that updates aligned with human feedback within PERSCO significantly improve perceived safety ( $$p<.001$$ p < . 001 ) relative to an adversarial adaptation condition, and that collective updates (where parameters are shared across users) further reduce feedback demands.