[論文レビュー] MeCo: Enhancing LLM-Empowered Multi-Robot Collaboration via Similar Task Memoization

Baiqing Wang, Helei Cui+4 more|arXiv (Cornell University)|Jan 28, 2026

Multimodal Machine Learning Applications被引用数 0

ひとこと要約

MeCo introduces a similarity-aware framework that caches and reuses solutions for similar multi-robot tasks, using a Similar Task Memoization strategy and a Similar Motion Planner (S-Planner) to reduce LLM calls, planning time, and token usage.

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.

研究の動機と目的

• Motivate reducing planning costs and token usage in LLM-powered multi-robot collaboration by exploiting task similarities.
• Develop a similarity testing mechanism to identify relevant past tasks for reuse.
• Design a Similar Motion Planner (S-Planner) to reuse plans from similar tasks and avoid full LLM replanning.
• Create MeCoBench to evaluate performance on similar-task collaboration scenarios.
• Provide an open-source implementation of MeCo.

提案手法

• Introduce a task cache that stores successful task plans and retrieves similar tasks for reuse.
• Define similarity criteria for low- and high-workspace-overlap tasks to select past tasks.
• Develop S-Planner to inherit and adapt plans from similar tasks; include back checking and collision checking for high-overlap tasks.
• Implement continuous planning to re-enter LLM planning from the failure step when S-Planner cannot complete a task.
• Incorporate selective caching with an LFU-inspired deduplication mechanism to manage cache size and task diversity.
• Extend RoCoBench to form MeCoBench for evaluating similar-task performance; compare against RoCo, Central Plan, HMAS-2, and ReAct.

実験結果

リサーチクエスチョン

• RQ1 How can task-level similarity be defined and detected to enable plan reuse in multi-robot collaboration?
• RQ2 To what extent can reusing similar-task plans reduce LLM invocations, planning time, and token consumption without sacrificing success rates?
• RQ3 Does the proposed S-Planner reliably adapt or reuse plans across varying workspace overlap scenarios?
• RQ4 What is the impact of cache size and deduplication strategy on MeCo’s performance across different task families?
• RQ5 Can continuous planning effectively bridge gaps when S-Planner cannot fully complete a task?

主な発見

• MeCo achieves around 30% improvement in task success rate over baselines in random and similar scenarios.
• MeCo reduces planning time by about 55% and token consumption by up to 70% compared with state-of-the-art baselines.
• For high-workspace-overlap tasks, larger task caches improve success rates and reduce tokens, but planning time exhibits a decrease-then-increase trend due to search overhead.
• For low-workspace-overlap tasks, larger caches generally improve success rates and decrease tokens with planning time stabilizing.
• In MeCoBench, MeCo consistently outperforms baselines in S1 (random) and S2 (totally similar), with performance approaching parity with baselines in S3 (totally different).
• A significant portion of gains comes from the similarity testing, S-Planner integration, and continuous planning components, as shown by ablation studies.