Enhancing Lifelong Multi-Agent Path-finding by Using Artificial Potential Fields

TL;DR

This paper investigates applying Artificial Potential Fields (APFs) to Multi-Agent Path Finding (MAPF) and Lifelong MAPF (LMAPF). It first shows that direct APFs are ineffective for offline MAPF, then integrates APFs into TA*, SIPPS, PIBT, and LaCAM to bias searches away from congested areas. In Lifelong MAPF, APF-augmented methods substantially boost system throughput, up to about sevenfold in some benchmarks, by reducing future conflicts as new goals arrive. The study provides a comprehensive parameter analysis and demonstrates that APFs offer practical benefits for online, dynamic tasking scenarios, while recommending careful tuning. Overall, the work highlights APFs as a viable congestion-avoidance tool in LMAPF but not in static MAPF, with meaningful implications for real-time autonomous systems.

Abstract

We explore the use of Artificial Potential Fields (APFs) to solve Multi-Agent Path Finding (MAPF) and Lifelong MAPF (LMAPF) problems. In MAPF, a team of agents must move to their goal locations without collisions, whereas in LMAPF, new goals are generated upon arrival. We propose methods for incorporating APFs in a range of MAPF algorithms, including Prioritized Planning, MAPF-LNS2, and Priority Inheritance with Backtracking (PIBT). Experimental results show that using APF is not beneficial for MAPF but yields up to a 7-fold increase in overall system throughput for LMAPF.

Figures (11)

• Figure 1: (a) An instance that cannot be solved by DAPF. The dotted lines depict a PrP solution. (b) Two agents solve LMAPF. The $X$ shapes represent goal locations. A solid blue line shows the direction of a chosen path for the blue agent. Dashed red lines represent alternative $k$-length paths for the orange agent. The agent will prefer the bottom path because of the APFs of the path of the blue agent.
• Figure 2: The orange circle and the orange square are the agent's start and goal locations, respectively. $x$ represents the cost of APFs. A$^*$ is executed. $g, h$ and $f$ are the components of A$^*$ nodes. (a) With $x\leq2$, an agent always picks a red path wherever $x$ is added. (b) With $x>2$, the agent picks a blue path if $x$ is added to $g$. Otherwise, if $x$ is added to $h$, it continues to choose a red path nonetheless.
• Figure 3: LMAPF: Average Throughput. Dashed lines - APF-enhanced; Solid lines - no APFs
• Figure 4: Parameter sensitivity analysis. The "L-PIBT" label designates the original PIBT algorithm that does not utilize APFs.
• Figure 5: The influence of parameters' values on APFs. (a) $w$ controls the strength; (b) $\gamma$ controls the rate of decay; (c) $d_{max}$ defines the radius of influence.