Anticipating Human Collision Avoidance Behavior for Safe Robot Reaction

摘要

For robots to effectively navigate in the presence of humans, they must safely leverage the human's perceived unwillingness to collide. Drawing on Viability Theory, we propose a novel approach to robustly anticipate human collision-avoiding behavior. We assume that rational humans try to optimally control their motion to avoid collision, but they are also prone to error, which makes their behavior suboptimal. We offer a robust control model which varies the level of optimality expected over time, assuming that humans may act unpredictably for a brief period of time, but their actions approach optimal collision-avoiding behavior as time progresses. We show how the proposed model can be used to produce a set of initial states for which a rational human will avoid collision. Further, we produce a robust policy which characterizes the set of control inputs expected by the human at any state. We illustrate our approach using two representative scenarios.