CoMuRoS: A LLM-Driven Hierarchical Architecture for Adaptive Multi-Robot Collaboration

Suraj Borate, Bhavish Rai B, Vipul Pardeshi, Madhu Vadali

Keywords: Task Planning, LLM, Multi-robot, Event-Driven, Human-robot Collaboration, Heterogeneous Team

TL;DR: CoMuRoS is an LLM-powered, event-driven task planning and execution framework that embodies a human-inspired manager–team model for multi-robot collaboration.

Abstract:

Dynamic environments require multi-robot teams to adapt to disruptive events by finding collaborative plans to achieve the goals. We introduce CoMuRoS (Collaborative Multi-Robot System)\cite{comuros}, an LLM-powered hierarchical architecture for heterogeneous robots that enables zero-shot task classification, allocation, and event-triggered replanning. A central task manager LLM processes natural language goals, categorizes tasks (independent, sequential, coordinated, infeasible) and assigns them based on capabilities and context. Local robot LLMs generate executable code from primitives, with perception classifying events to initiate replanning for recovery or human aid.
Hardware tests validate adaptation to failures (e.g.,collaborative object recovery: 9/10 success and emergent human-multi-robot collaboration (5/5)). Simulations demonstrate architecture's applicability for understanding and responding to user intentions in health care and disaster relief scenario. A 22-scenario benchmark demonstrates generalization (correctness: 0.91 with Grok-3). CoMuRoS advances reasoning and planning in dynamic environments by blending centralized deliberation with decentralized execution.

Submission Number: 28