Abstract

We consider a team of robots with heterogeneous capabilities tasked with accomplishing a set of tasks in the environment. In such a setting, we present a learning-based approach to resilient coalition formation for robots operating under environmental disturbances. Towards this end, we leverage human expert-demonstrations, which record the coalition-task pairs generated by experts to satisfy task requirements in the presence of environmental disturbances and degraded robot operations. We design a task monitor, which quantifies the performance of robots in tasks to detect which capabilities of the robots are compromised—and learn a mapping from the failure information to the reallocations generated by the users. Preliminary results show promise in automatically synthesizing heterogeneous coalitions based on collected expert data.