Scaling Lifelong Multi-Agent Path Finding to More Realistic Settings: Research Challenges and Opportunities

TL;DR

This work tackles scaling Lifelong MAPF to very large, dense agent populations under tight time budgets. It introduces Windowed Parallel PIBT-LNS (WPPL), a fast, scalable approach that combines windowed planning with PIBT-LNS refinement and parallel execution, achieving strong throughput in the 2023 LRR competition. The paper identifies three core challenges—limited planning time, traffic congestion/myopia, and gaps between theoretical LMAPF models and real-world systems—and proposes concrete directions: improved rule-based or MARL methods, guidance graphs, and agent-disabling strategies to mitigate congestion, plus rotation-inclusive modeling and realistic uncertainty considerations. The findings show WPPL achieving top performance and provide a roadmap for future research to bridge theory and practice in real-world, large-scale LMAPF applications.

Abstract

Multi-Agent Path Finding (MAPF) is the problem of moving multiple agents from starts to goals without collisions. Lifelong MAPF (LMAPF) extends MAPF by continuously assigning new goals to agents. We present our winning approach to the 2023 League of Robot Runners LMAPF competition, which leads us to several interesting research challenges and future directions. In this paper, we outline three main research challenges. The first challenge is to search for high-quality LMAPF solutions within a limited planning time (e.g., 1s per step) for a large number of agents (e.g., 10,000) or extremely high agent density (e.g., 97.7%). We present future directions such as developing more competitive rule-based and anytime MAPF algorithms and parallelizing state-of-the-art MAPF algorithms. The second challenge is to alleviate congestion and the effect of myopic behaviors in LMAPF algorithms. We present future directions, such as developing moving guidance and traffic rules to reduce congestion, incorporating future prediction and real-time search, and determining the optimal agent number. The third challenge is to bridge the gaps between the LMAPF models used in the literature and real-world applications. We present future directions, such as dealing with more realistic kinodynamic models, execution uncertainty, and evolving systems.

Figures (8)

• Figure 1: Three challenging maps in the LRR competition with white vertices and black obstacles. (a) is a random map of size 32 by 32 with very high agent density (800 agents out of 819 vertices ). The colorful triangles are agents, with short black lines indicating orientations. (b) is a large sortation center map of size 140 by 500 with 54,230 vertices and 10,000 agents. (c) is a city map representing a part of Paris.
• Figure 2: Windowed Parallel PIBT-LNS (WPPL). For each windowed MAPF instance, we run PIBT $w$ steps to get an initial plan $a_{1:w}$ and use Parallel MAPF-LNS to refine it. We then execute the first $h$ actions of the refined plan $a_{1:h}$, updating the MAPF instance and reusing the rest of actions $a_{h+1:w}$ in the next iteration of PIBT.
• Figure 3: The anytime behavior of WPPL. For better illustration, we apply guidance graphs (introduced later in \ref{['section:myopic']}), use only a single thread, and normalize throughput by the number of agents in this experiment.
• Figure 4: The effect of Parallelizing LNS. For better illustration, we apply guidance graphs (introduced later in \ref{['section:myopic']}) and normalize throughput by the number of agents in this experiment. This figure uses the same legend as \ref{['fig:anytime']}.
• Figure 5: Comparison of with and without guidance graph in the Warehouse 8000 instance. The heatmap shows the wait action usage (the number of steps agents wait in each vertex). Red denotes areas of high congestion.

Theorems & Definitions (3)

• Definition 1: MAPF
• Definition 2: LMAPF
• Definition 3: MAPF with Rotations (MAPF-R)