Perception-Based Rewards for Robotic Shepherd Teams Maneuvering Split Flocks

Anonymous

Keywords: Motion Planning, Robot Shepherding

TL;DR: Coordinated active agents (shepherds) influencing passive reactive agents (sheep) can be taught to perform complex collection tasks while achieving shepherding goals.

Abstract: Underactuated system tasks, like shepherding passive agents using active coordinated robotic agent teams, require quick reactions and consistent perception and control. A recent learning-based solution demonstrated the agility required for such a task, but only accounted for single cohesive flocks. Non-contiguous flocks, on the other hand, can diffuse if not handled in a timely fashion. We address the disjoint flock case by defining novel reward schemes, based on the shepherds' visual observations. We show that policies trained on these rewards succeed at shepherding disjoint and fractious flocks to a goal region in a motion-efficient manner, and provide comparisons to state of the art learning-based and heuristic methods.