LLMDR: LLM-Driven Deadlock Detection and Resolution in Multi-Agent Pathfinding (2503.00717v1)

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 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

Insights into LLM-Driven Deadlock Detection and Resolution in Multi-Agent Pathfinding

The paper "LLMDR: LLM-Driven Deadlock Detection and Resolution in Multi-Agent Pathfinding" presents a novel approach for addressing a critical issue in Multi-Agent Pathfinding (MAPF): deadlocks. Within multi-agent systems, MAPF is known for its complexity and tendency to result in deadlocks, particularly as the number of agents and environmental complexity increase. The proposed strategy, LLMDR, leverages LLMs to enhance the performance of learned MAPF models by detecting and resolving deadlocks.

Overview of LLMDR Approach

LLMDR integrates the strengths of LLMs—known for their robust decision-making capabilities—and the efficiency of prioritized planning augmented through the Priority Inheritance with Backtracking (PIBT) algorithm. The framework operates through two primary phases: deadlock detection and deadlock resolution.

1. Deadlock Detection: The LLM examines a detection window of agent movements to identify patterns indicative of deadlocks, such as immobility or erratic wandering. Agents are grouped based on proximity and goal achievement status, and LLM strategies classify these groups as needing resolution through "leader" or "radiation" methods.
2. Deadlock Resolution: Once deadlocks are detected, the system moves to resolution, where the LLM sets strategies that prioritize actions for agents—either promoting directional movement for "leaders" or dispersal from congested zones for "radiate" agents. These strategies dictate the agent and action priorities for PIBT, allowing for efficient short-term deadlock resolution.

Experimental Evaluation

The research evaluates LLMDR using several MAPF benchmark maps, incorporating different agent numbers to assess scalability and effectiveness of deadlock resolution across diverse scenarios. LLMDR was tested in combination with various base models, like DCC, DHC, and SACHA, showing significant improvements in success rates and reduction in episode lengths, especially in complex, agent-dense scenarios. This improvement suggests that deadlocks are a primary failure mode for many MAPF models, and effective resolution dramatically enhances system performance.

LLM Impact and Hyperparameter Considerations

The paper highlights the influence of LLM capabilities by contrasting gpt-3.5-turbo and gpt-4o within the LLMDR framework. The results demonstrate that higher capability LLMs offer superior performance, inferred by shorter episode lengths and higher success rates, especially in challenging environments. Additionally, a hyperparameter analysis underscores the importance of calibrating detection window and execution plan lengths, balancing effective intervention with computational efficiency.

Implications and Future Directions

The integration of LLMs into MAPF optimization signals a significant shift in addressing the scalability issues inherent in traditional methods. However, computational overhead remains a concern, which restricts the use of LLMDR in low-resource or real-time applications. Future explorations could focus on optimizing the architecture of LLMs or employing hybrid models to mitigate these constraints and further enhance the applicability of LLMDR.

In conclusion, the paper successfully demonstrates the potential of LLMs for improving the robustness and efficiency of multi-agent pathfinding by effectively resolving deadlocks, thus broadening the horizon for their application in complex, dynamic environments.