PhD Dissertation Proposal: Williard Joshua Jose, Adaptive Multi-Agent Learning for Multi-Domain Collaborative Autonomy
Content
Speaker:
Abstract:
Multi-robot teams require coordinated decision-making across different domains in order to complete complex tasks safely and efficiently. Achieving this coordination involves solving interconnected problems ranging from high-level decision-making for task scheduling, to tightly-coupled physical coordination and individual robot trajectory planning. Because collaborative robots operate in diverse environments with varying heterogeneous team compositions, kinematic constraints, and runtime objectives, individual learning policies are difficult to generalize across all scenarios. To address this, this dissertation aims to explore the different learning approaches on three domains: high-level decision-making for multi-robot systems, tightly-coupled physical coordination of multi-robot execution, and execution-time guidance of collaborative policies. For the high-level decision making, I investigate graph-based and imitation learning models to enable heterogeneous robots to dynamically form subteams for task scheduling. For the tightly-coupled physical coordination, I apply constrained diffusion and reinforcement learning to couple team-level trajectory generation with individual robot constraints for multi-quadruped collaborative transportation. For execution-time guidance, I formulate multi-agent reinforcement learning with guided flow policies to adapt to dynamically changing execution-time objectives. Finally, in this dissertation, I discuss extensions to improve competence-aware scheduling for high-level decision-making, bilevel learning for tightly-coupled physical coordination, and expressive execution-time guidance through collaborative lookahead flow guidance.
Advisor:
Hao Zhang