Robust Execution of Plans for Human-Robot Teams

Abstract

Humans and robots working together can efficiently complete tasks that are very difficult for either to accomplish alone. To collaborate fluidly, robots must recognize the humans' intentions and adapt to their actions appropriately. Pike is an online executive introduced previously in the literature that unifies intent recognition and plan adaptation for temporally flexible plans with choice. While successful at coordinating human-robot teams, Pike had limited robustness to temporal uncertainty about the durations of actions. This paper presents two extensions to Pike that make it much more robust to temporal uncertainty. First, we extend Pike to handle uncontrollable action durations by enforcing strong temporal controllability. We accomplish this by generalizing standard strong controllability algorithms for STNUs to plans with choice. Second, in case a realized duration exceeds even the specified bounds and makes the entire plan infeasible, we attempt to intelligently negotiate with a human to relax some of the temporal constraints and restore feasibility, rather than immediately failing and halting execution. This negotiation is guided by a state-of-the-art conflict directed relaxation algorithm, which has previously only been used offline.