Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLMDR: LLM-Driven Deadlock Detection and Resolution in Multi-Agent Pathfinding

Seungbae Seo, Junghwan Kim, Minjeong Shin, Bongwon Suh

TL;DR

LLMDR addresses deadlocks in learning-based MAPF by coupling LLM-driven deadlock detection with LLM-guided resolution, executed through PIBT to produce collision-free 1-step actions. The method detects deadlocks within a detection window, groups proximate deadlocks, and assigns leader or radiation strategies to guide priorities, iteratively resolving conflicts as the plan unfolds. Experiments on multiple MAPF benchmarks show that LLMDR improves success rates and reduces episode lengths across several base models, with stronger gains in deadlock-prone, larger-agent scenarios, and when using higher-capability LLMs (e.g., GPT-4o). While effective, the approach incurs substantial computational cost from LLM inference, indicating future work in efficiency and hybrid designs for real-time applicability.

Abstract

Multi-Agent Pathfinding (MAPF) is a core challenge in multi-agent systems. Existing learning-based MAPF methods often struggle with scalability, particularly when addressing complex scenarios that are prone to deadlocks. To address these challenges, we introduce LLMDR (LLM-Driven Deadlock Detection and Resolution), an approach designed to resolve deadlocks and improve the performance of learnt MAPF models. LLMDR integrates the inference capabilities of large language models (LLMs) with learnt MAPF models and prioritized planning, enabling it to detect deadlocks and provide customized resolution strategies. We evaluate LLMDR on standard MAPF benchmark maps with varying agent numbers, measuring its performance when combined with several base models. The results demonstrate that LLMDR improves the performance of learnt MAPF models, particularly in deadlock-prone scenarios, with notable improvements in success rates. These findings show the potential of integrating LLMs to improve the scalability of learning-based MAPF methods. The source code for LLMDR is available at: https://github.com/ssbacc/llmdr-dhc

LLMDR: LLM-Driven Deadlock Detection and Resolution in Multi-Agent Pathfinding

TL;DR

LLMDR addresses deadlocks in learning-based MAPF by coupling LLM-driven deadlock detection with LLM-guided resolution, executed through PIBT to produce collision-free 1-step actions. The method detects deadlocks within a detection window, groups proximate deadlocks, and assigns leader or radiation strategies to guide priorities, iteratively resolving conflicts as the plan unfolds. Experiments on multiple MAPF benchmarks show that LLMDR improves success rates and reduces episode lengths across several base models, with stronger gains in deadlock-prone, larger-agent scenarios, and when using higher-capability LLMs (e.g., GPT-4o). While effective, the approach incurs substantial computational cost from LLM inference, indicating future work in efficiency and hybrid designs for real-time applicability.

Abstract

Multi-Agent Pathfinding (MAPF) is a core challenge in multi-agent systems. Existing learning-based MAPF methods often struggle with scalability, particularly when addressing complex scenarios that are prone to deadlocks. To address these challenges, we introduce LLMDR (LLM-Driven Deadlock Detection and Resolution), an approach designed to resolve deadlocks and improve the performance of learnt MAPF models. LLMDR integrates the inference capabilities of large language models (LLMs) with learnt MAPF models and prioritized planning, enabling it to detect deadlocks and provide customized resolution strategies. We evaluate LLMDR on standard MAPF benchmark maps with varying agent numbers, measuring its performance when combined with several base models. The results demonstrate that LLMDR improves the performance of learnt MAPF models, particularly in deadlock-prone scenarios, with notable improvements in success rates. These findings show the potential of integrating LLMs to improve the scalability of learning-based MAPF methods. The source code for LLMDR is available at: https://github.com/ssbacc/llmdr-dhc

Paper Structure

This paper contains 16 sections, 2 figures, 3 tables, 2 algorithms.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Learning-Based MAPF Methods
  4. LLM for Planning
  5. Preliminaries
  6. Multi-Agent Pathfinding
  7. Prioritized Planning and PIBT
  8. LLMDR: LLM-Driven Deadlock Detection and Resolution
  9. Deadlock Detection by LLM
  10. Deadlock Resolution Using LLM Strategies with PIBT
  11. Experimental Results and Analysis
  12. Experiment Setup
  13. LLMDR Performance with Various Base Models
  14. Impact of Language Model Capability on LLMDR
  15. Hyperparameter Analysis
  16. ...and 1 more sections

Figures (2)

  • Figure 1: Overview of LLMDR (LLM-Driven Deadlock Detection and Resolution). When a deadlock is detected in the base model simulation, manual actions are iterated to resolve the deadlock. If no deadlock is detected, the base model simulation is directly applied to the environment.
  • Figure 2: Detailed implementation of LLMDR (LLM-Driven Deadlock Detection and Resolution). The figure illustrates the process where agents in deadlock situations are detected by the LLM, and the action sets for each agent to resolve the deadlocks are obtained using the strategies proposed by the LLM along with PIBT algorithm. PIBT is used to derive collision-free actions from strategies generated by the LLM, thus enabling cooperative behavior among agents to resolve deadlocks.