MeCo: Enhancing LLM-Empowered Multi-Robot Collaboration via Similar Task Memoization

TL;DR

MeCo tackles the inefficiency of re-planning in LLM-driven multi-robot collaboration by introducing similarity-aware memoization. It defines task similarity across high- and low-workspace-overlap scenarios, introduces a Similar Motion Planner (S-Planner) and a continuous planning module, and uses a selective LFU-based cache to reuse prior task plans. Across MeCoBench experiments, MeCo significantly improves task success and reduces planning time and token usage, particularly when tasks are similar, while maintaining competitive performance in dissimilar cases. The work advances practical, scalable LLM-assisted multi-robot planning by exploiting inter-task relations to curb runtime costs and improve robustness, with open-source tooling and a dedicated benchmark for similar-task scenarios.

Abstract

Multi-robot systems have been widely deployed in real-world applications, providing significant improvements in efficiency and reductions in labor costs. However, most existing multi-robot collaboration methods rely on extensive task-specific training, which limits their adaptability to new or diverse scenarios. Recent research leverages the language understanding and reasoning capabilities of large language models (LLMs) to enable more flexible collaboration without specialized training. Yet, current LLM-empowered approaches remain inefficient: when confronted with identical or similar tasks, they must replan from scratch because they omit task-level similarities. To address this limitation, we propose MeCo, a similarity-aware multi-robot collaboration framework that applies the principle of ``cache and reuse'' (a.k.a., memoization) to reduce redundant computation. Unlike simple task repetition, identifying and reusing solutions for similar but not identical tasks is far more challenging, particularly in multi-robot settings. To this end, MeCo introduces a new similarity testing method that retrieves previously solved tasks with high relevance, enabling effective plan reuse without re-invoking LLMs. Furthermore, we present MeCoBench, the first benchmark designed to evaluate performance on similar-task collaboration scenarios. Experimental results show that MeCo substantially reduces planning costs and improves success rates compared with state-of-the-art approaches.

Figures (15)

• Figure 1: Workflow of MeCo. The arrow $\rightarrow$ shows the LLM-empowered planning; the arrow $\Rightarrow$ shows the MeCo extensions.
• Figure 2: The value of $\alpha_i$ affects the success rate of planning the current task based on similar tasks. We test the "Pick" and "Place" subtasks of the Pack Grocery and Move Rope tasks under different intervals, with each setting tested 100 times on average.
• Figure 3: Workflow of S-Planner in the high-workspace-overlap mode. The left side illustrates how tasks are decomposed and assigned at the high level, while the right side shows how motion planning is conducted at the low level. The blue curves represent the reference trajectory, the red points mark the target positions in the similar task; 'Alice' and 'Bob' refer to robots.
• Figure 4: Performance vs. cache task count. As the number of tasks stored in the cache increases, we observe changes in the average success rate, planning time, and token consumption of MeCo across different tasks. For each task type, we randomly store tasks planned by LLM-empowered methods in the task cache and evaluate MeCo averaged over 30 random seeds.
• Figure 5: The performance of MeCo. We evaluate on MeCoBench across three scenarios: random (S1), totally similar (S2), and totally different (S3), using four baselines for comparison. For each task, we report success rate, planning time, and token consumption averaged over 30 random seeds. The results on Sweep Floor and Arrange Cabinet are provided in the Appendix \ref{['A1']}.