Probabilistically Safe Robot Planning with Confidence-Based Human\n Predictions

Abstract

In order to safely operate around humans, robots can employ predictive models\nof human motion. Unfortunately, these models cannot capture the full complexity\nof human behavior and necessarily introduce simplifying assumptions. As a\nresult, predictions may degrade whenever the observed human behavior departs\nfrom the assumed structure, which can have negative implications for safety. In\nthis paper, we observe that how "rational" human actions appear under a\nparticular model can be viewed as an indicator of that model's ability to\ndescribe the human's current motion. By reasoning about this model confidence\nin a real-time Bayesian framework, we show that the robot can very quickly\nmodulate its predictions to become more uncertain when the model performs\npoorly. Building on recent work in provably-safe trajectory planning, we\nleverage these confidence-aware human motion predictions to generate assured\nautonomous robot motion. Our new analysis combines worst-case tracking error\nguarantees for the physical robot with probabilistic time-varying human\npredictions, yielding a quantitative, probabilistic safety certificate. We\ndemonstrate our approach with a quadcopter navigating around a human.\n