CAREER: Robust and Reliable Multiagent Scheduling under Uncertainty

Harvey Mudd College, Claremont CA

Investigators

Abstract

Planning is important for autonomous systems, and planning for the real world typically involves reasoning about uncertainty in perception, action, and how the environment will react to actions of the agents. This work will improve the robustness and reliability or plans in applications such as autonomous driving, automated warehousing, and personal robots by addressing limitations in how current planning systems handle real-world scheduling uncertainty. The research will explore fundamental questions such as: What makes a plan good? How good is it? How can we make it better? And, how should we adjust plans when faced with uncertainty? The research will produce automated planning and scheduling techniques that can robustly adapt to real-world, uncertain interactions with physical environments and teammates. The project will also develop and share curricula for two new undergraduate courses in robotics that highlight this research and will host a national AI Predoctoral Workshop that aims to broaden the pipeline of under-represented students into AI graduate research. This work addresses a fundamental gap that currently exists between AI temporal planning theory and the execution of such plans, in practice. The specific research objectives are to (1) equip agents with more accurate plan representations by learning models of temporal uncertainty; (2) develop more useful measures of plan quality by introducing novel robustness and reliability metrics that predict the prospects of successful execution; (3) design more resilient scheduling methods by devising new online and offline heuristics that hedge against uncertainty; (4) construct more durable multi-agent coordination protocols by creating new decentralized algorithms for decoupling agents' schedules to allow robust, independent execution; and (5) demonstrate the efficacy of these ideas in the real world by evaluating on a diverse corpus of multi-robot benchmark problems. If successful, this research will significantly improve the efficiency and usefulness of existing planning techniques in real-world settings and reframe how researchers analyze temporal plans.