Leveraging Adaptive Group Negotiation for Heterogeneous Multi-Robot Collaboration with Large Language Models

TL;DR

CLiMRS tackles the challenge of coordinating heterogeneous robots over long horizons by employing an adaptive group negotiation framework in which each robot is paired with an independent LLM agent. A dynamic grouping planner forms subgroups, local subgroup managers generate task-level commands, and execution coupled with a feedback-rich environment updates the context memory for subsequent planning, enabling robust long-horizon performance. The authors introduce CLiMBench, a physics-based benchmark with diverse robot types and long-horizon assembly tasks to test planning-perception interaction under uncertainty. Across COHERENT and CLiMBench, CLiMRS achieves higher efficiency and reliability than strong baselines, with ablations demonstrating the essential roles of dialogue history, execution feedback, and grouping. The work suggests that human-inspired dynamic grouping and negotiation principles can substantially improve heterogeneous multi-robot collaboration in realistic settings, with potential for scalable deployment and safer, more robust autonomous teams.

Abstract

Multi-robot collaboration tasks often require heterogeneous robots to work together over long horizons under spatial constraints and environmental uncertainties. Although Large Language Models (LLMs) excel at reasoning and planning, their potential for coordinated control has not been fully explored. Inspired by human teamwork, we present CLiMRS (Cooperative Large-Language-Model-Driven Heterogeneous Multi-Robot System), an adaptive group negotiation framework among LLMs for multi-robot collaboration. This framework pairs each robot with an LLM agent and dynamically forms subgroups through a general proposal planner. Within each subgroup, a subgroup manager leads perception-driven multi-LLM discussions to get commands for actions. Feedback is provided by both robot execution outcomes and environment changes. This grouping-planning-execution-feedback loop enables efficient planning and robust execution. To evaluate these capabilities, we introduce CLiMBench, a heterogeneous multi-robot benchmark of challenging assembly tasks. Our experiments show that CLiMRS surpasses the best baseline, achieving over 40% higher efficiency on complex tasks without sacrificing success on simpler ones. Overall, our results demonstrate that leveraging human-inspired group formation and negotiation principles significantly enhances the efficiency of heterogeneous multi-robot collaboration. Our code is available here: https://github.com/song-siqi/CLiMRS.

Figures (12)

• Figure 1: Overview. We present CLiMRS, an adaptive group negotiation framework among multiple LLMs that enables multi-robot collaboration through a grouping–planning–execution–feedback loop, and CLiMBench, a heterogeneous multi-robot collaboration benchmark in simulation with challenging long-horizon assembly tasks.
• Figure 2: CLiMRS Framework. To employ our grouping–planning–execution–feedback cycle, CLiMRS comprises (a) a general proposal planner that forms dynamic agent subgroups, (b) multiple subgroup managers for local agent commands, (c) multiple agent executors for robot skills and agent feedback, (d) a simulation environment for environment interaction and feedback, and (e) a context memory module for all dialogue context and feedback.
• Figure 3: Our Benchmark. CLiMBench is a heterogeneous multi-robot collaboration benchmark, featuring multi-agent robots with diverse skills, enabling collaboration on complex assembly tasks of varying difficulty levels.
• Figure 4: Ablation Study on AMP Domain Randomization. Results show that Domain Randomization helps the humanoid agent to obtain a stable and reusable carrying skill in CLiMBench.
• Figure 5: Single Agent Prompt Templates. This template encodes the agent’s name, unique identifier, and action specification to ensure structured and unambiguous command interpretation for an agent executor.