Adaptive Risk-Based Replanning For Human-Aware Multi-Robot Task Allocation With Local Perception

In this paper, we propose an adaptive risk-based replanning strategy in the context of multi-robot task allocation for dealing with limitations of local perception and unpredicted human behavior. Our replanning method is based on the variations of social risk and human motion prediction uncertainty. The performance of our method is studied through an extensive suite of experiments of increasing complexity. Results obtained using both a high-fidelity simulator and real robots confirm that this strategy outperforms a non-adaptive replanning strategy in all cases with respect to the chosen social metrics. The overall performance of the team depends firstly on its replanning strategy, and secondly on the available information about the humans. Although an adaptive replanning strategy with global perception leads to the best performance, it is computationally expensive and infeasible in some real applications. Local perception shows comparable results as long as updates of relevant human poses affecting a task's risk are available within the execution time of that task. Conversely, the non-adaptive replanning strategy is shown to have degraded results with global perception as decisions in this case can be based on outdated information that lead to invalid plans.