Graph Attention-Guided Search for Dense Multi-Agent Pathfinding

TL;DR

LaGAT presents a hybrid MAPF framework that embeds an enhanced neural heuristic MAGAT+ into the LaCAM search algorithm, augmented with a deadlock-detection mechanism to handle imperfect neural guidance. MAGAT+ is pre-trained on large-scale MAPF instances and map-wise fine-tuned to a target map, enabling long-horizon coordination in dense environments; neural guidance is combined with PIBT-based search, with deadlock detection overriding guidance when necessary to preserve safety and completeness. Evaluations show LaGAT delivers higher-quality initial and final solutions in dense MAPF settings compared to both LaCAM and MAPF-GPT, while maintaining practical runtimes, and demonstrate that pre-training, fine-tuning, deadlock detection, and search are all critical to performance. The work highlights a plug-and-play, complete neuro-guided search paradigm that redefines the speed–quality Pareto frontier for dense multi-agent coordination and suggests broad applicability to related domains.

Abstract

Finding near-optimal solutions for dense multi-agent pathfinding (MAPF) problems in real-time remains challenging even for state-of-the-art planners. To this end, we develop a hybrid framework that integrates a learned heuristic derived from MAGAT, a neural MAPF policy with a graph attention scheme, into a leading search-based algorithm, LaCAM. While prior work has explored learning-guided search in MAPF, such methods have historically underperformed. In contrast, our approach, termed LaGAT, outperforms both purely search-based and purely learning-based methods in dense scenarios. This is achieved through an enhanced MAGAT architecture, a pre-train-then-fine-tune strategy on maps of interest, and a deadlock detection scheme to account for imperfect neural guidance. Our results demonstrate that, when carefully designed, hybrid search offers a powerful solution for tightly coupled, challenging multi-agent coordination problems.

Figures (11)

• Figure 1: Pareto frontier consisting of state-of-the-art algorithms in dense MAPF instances.
• Figure 2: Overview of LaGAT. We utilize 21K instances for pre-training (a, b) and 1K instances for fine-tuning (c). The resulting policy MAGAT+ is embedded into LaCAM search over configurations (d). When a deadlock is detected for a specific agent, the search performs a rollback, and the neural guidance is selectively overridden by the default, non-learning PIBT mechanism.
• Figure 3: Evaluation on neural policies by success rate and runtime. Model sizes are listed in the bottom right corner.
• Figure 4: Map-wise evaluation. Runtime and sum-of-costs scores are average within solved instances for each method. Transparent regions represent 95% confidence intervals. SoC is normalized by its trivial lower bound $\sum_{i\in A}\mathsf{dist}\xspace(s_i, g_i)$. MAPF-GPT is significantly slower than the other methods and falls outside the plotted runtime range. For reference, we also display the results of another leading search-based solver, LNS2 li2022mapf. With a timeout of 30s, LNS2 fails in dense setups (e.g., 0% in Dense Warehouse with 128 agents), so we only plot LNS2 data points with a success rate higher than 80%.
• Figure 5: Instance-wise sum-of-costs comparison between lacam3 and LaGAT across all instances included in \ref{['fig:main-eval']}. Points in the upper-left triangular region indicate where LaGAT outperforms lacam3.

Theorems & Definitions (2)

• Theorem 1
• Theorem 1