Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Local Guidance for Configuration-Based Multi-Agent Pathfinding

Tomoki Arita, Keisuke Okumura

TL;DR

Local guidance addresses congestion in MAPF by providing spatiotemporal cues around each agent, rather than relying on global congestion models. The method integrates local guidance into LaCAM via windowed planning, a lexicographic cost with collision penalties, and a set of practical design choices (initialization, plan iteration, agent ordering, and optional global guidance). Empirical results on benchmarks with up to 10,000 agents show substantial flowtime improvements (up to ~50% in some cases) and competitiveness with state-of-the-art methods like LNS2, often outperforming global guidance. The approach achieves a practical balance between solution quality and runtime, and code is publicly available.

Abstract

Guidance is an emerging concept that improves the empirical performance of real-time, sub-optimal multi-agent pathfinding (MAPF) methods. It offers additional information to MAPF algorithms to mitigate congestion on a global scale by considering the collective behavior of all agents across the entire workspace. This global perspective helps reduce agents' waiting times, thereby improving overall coordination efficiency. In contrast, this study explores an alternative approach: providing local guidance in the vicinity of each agent. While such localized methods involve recomputation as agents move and may appear computationally demanding, we empirically demonstrate that supplying informative spatiotemporal cues to the planner can significantly improve solution quality without exceeding a moderate time budget. When applied to LaCAM, a leading configuration-based solver, this form of guidance establishes a new performance frontier for MAPF.

Local Guidance for Configuration-Based Multi-Agent Pathfinding

TL;DR

Local guidance addresses congestion in MAPF by providing spatiotemporal cues around each agent, rather than relying on global congestion models. The method integrates local guidance into LaCAM via windowed planning, a lexicographic cost with collision penalties, and a set of practical design choices (initialization, plan iteration, agent ordering, and optional global guidance). Empirical results on benchmarks with up to 10,000 agents show substantial flowtime improvements (up to ~50% in some cases) and competitiveness with state-of-the-art methods like LNS2, often outperforming global guidance. The approach achieves a practical balance between solution quality and runtime, and code is publicly available.

Abstract

Guidance is an emerging concept that improves the empirical performance of real-time, sub-optimal multi-agent pathfinding (MAPF) methods. It offers additional information to MAPF algorithms to mitigate congestion on a global scale by considering the collective behavior of all agents across the entire workspace. This global perspective helps reduce agents' waiting times, thereby improving overall coordination efficiency. In contrast, this study explores an alternative approach: providing local guidance in the vicinity of each agent. While such localized methods involve recomputation as agents move and may appear computationally demanding, we empirically demonstrate that supplying informative spatiotemporal cues to the planner can significantly improve solution quality without exceeding a moderate time budget. When applied to LaCAM, a leading configuration-based solver, this form of guidance establishes a new performance frontier for MAPF.
Paper Structure (29 sections, 3 equations, 11 figures, 2 algorithms)

This paper contains 29 sections, 3 equations, 11 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries and Related Work
  3. Problem Definition.
  4. LaCAM
  5. How LaCAM Works.
  6. Configuration Generator
  7. PIBT
  8. Preference Construction in PIBT
  9. Guidance
  10. Method
  11. Exploitation.
  12. Cost and Path Planning (\ref{['algo:guidance:pathfinding']}).
  13. Plan Iteration (\ref{['algo:guidance:termination']}).
  14. Initialization (\ref{['algo:guidance:init']}).
  15. Agent Order (\ref{['algo:guidance:order']}).
  16. ...and 14 more sections

Figures (11)

  • Figure 1: Concept of global and local guidance with LaCAM. Goals for agents are marked with colored boxes. The gray-shaded regions correspond to configuration generation.
  • Figure 2: Vertex usage in LaCAM solutions, where congestion is implied by warm colors. The same start-goal instance comprising 1,000 agents on maze-128-128-10 is used across different solvers. The cost improvement percentage is shown at the top, with LaCAM as the baseline. The bottom shows the distribution among 14,818 vertices of how many times each vertex is used by any agent. In the middle, area-A shows the effect of global guidance, while area-B for local guidance.
  • Figure 3: Performance evaluation of suboptimal MAPF methods, assessed by flowtime and runtime. Top: Instance-wise comparison. From the MAPF benchmark, we extracted a total of 644 instances across five agent scales $\{200, 400, 600, 800, 1000\}$, selecting five instances per setting from 32 different grids. "cost reduction" is calculated using the LaCAM scores as a reference. The instances are sorted by the LG scores. Bottom: Map-wise analysis with varying numbers of agents from $200$ to $1000$. Each subfigure reports the average over the solved instances among 25 cases per setting. Note that LaCAM-based solvers successfully solved all instances within the 30s time limit, whereas LNS2 failed on den312d when $|A| \in \{800, 1000\}$. Flowtime scores are normalized by a trivial lower bound $\sum_{i\in A}\text{dist}(s_i, g_i)$. The number of vertices for each grid is shown with parentheses.
  • Figure 4: Scalability test on warehouse-20-40-10-2-2.
  • Figure 5: Effect of the collision cost $\alpha$ and window size $w$. "cost reduction" represents the ratio from LaCAM.
  • ...and 6 more figures