Co-chairs: Kira Barton and Maani Ghaffari Jadidi

Abstract:
With recent advancements in sensing and control, robots have become capable of an ever-growing range of single-agent tasks. On the other hand, there is still a lack of coordination frameworks that leverage the full capability of the single agents during complex multi-agent tasks. This dissertation tries to bridge the gap between single and multi-agent tasks through the development of a hierarchical multi-agent feedback planning system. This framework contains several key components: a task allocation planner responsible for forming agent teams and planning routes and schedules; a trajectory planner that ensures coordinated movements among agents; and a learning model that estimates agent capabilities and task requirements to provide feedback for updating the plans. It runs in pre-execution and real-time and applies both proactive and reactive methods to generate robust plans under uncertainty. The framework scales to hundreds of agents and generalizes to multiple problem types, including capture-the-flag games, robotic service allocation, multi-robot tour guiding, exploring and block picking, and cooperative manipulation.

Zoom passcode: bofuphd