Traffic Flow Optimisation for Lifelong Multi-Agent Path Finding

TL;DR

This work addresses MAPF scalability by introducing congestion-aware guide paths that anticipate traffic-induced delays. The authors extend PIBT and LaCAM* with two-part edge costs and congestion-driven guide heuristics, forming a Guided PIBT and guided LaCAM* framework. Empirical results demonstrate substantial throughput gains for lifelong MAPF with thousands of agents and improved solution quality for one-shot MAPF, though initialization and map-density tradeoffs exist. Overall, the approach enables scalable coordination for large robotic fleets and motivates further refinement of congestion-cost models in MAPF.

Abstract

Multi-Agent Path Finding (MAPF) is a fundamental problem in robotics that asks us to compute collision-free paths for a team of agents, all moving across a shared map. Although many works appear on this topic, all current algorithms struggle as the number of agents grows. The principal reason is that existing approaches typically plan free-flow optimal paths, which creates congestion. To tackle this issue, we propose a new approach for MAPF where agents are guided to their destination by following congestion-avoiding paths. We evaluate the idea in two large-scale settings: one-shot MAPF, where each agent has a single destination, and lifelong MAPF, where agents are continuously assigned new destinations. Empirically, we report large improvements in solution quality for one-short MAPF and in overall throughput for lifelong MAPF.

Figures (5)

• Figure 1: We show a small MAPF problem with 4 agents. Dashed (blue) lines indicate individually optimal paths from each $s_i$ to each $g_i$. Solid (purple) lines indicate congestion-aware individually-optimal paths.
• Figure 2: In this example high-priority agents enter a corridor just before a set of lower-priority agents exit, the number on each agent indicating its priority. Each time this occurs there is a large increase in the objective function value.
• Figure 3: A small figure with a guide path (the blue dashed line) and marked heuristic values.
• Figure 4: Lifelong MAPF. Average throughput (top) and response time in second (bottom). Shaded regions show standard deviation.
• Figure 5: Lifelong MAPF. Average tasks finished (top) and average response time (in the range 0-1 second; bottom). 24 instances, with 10,000 agents each. PIBT and GP-R1000 incur setup costs of 6.26 and 6.98 seconds to compute heuristic data.

Theorems & Definitions (1)

• Example 1