Caching-Augmented Lifelong Multi-Agent Path Finding

TL;DR

CAL-MAPF introduces a cache-augmented mechanism for Lifelong Multi-Agent Path Finding in warehouses by adding a Cache Grid near unloading ports, a cache locking system, and an external Task Assigner. The framework enables cache hits to reduce travel, while the lock mechanism prevents race conditions among concurrent agents. Through experiments with multiple input-task distributions and cache policies, the authors show that cache hit rate and traffic smoothness are critical factors driving performance, with gains most evident under favorable distributions and map configurations. The work highlights practical potential for throughput improvements in ongoing, dynamic warehouse operations and points to future improvements via smarter task assignment and more advanced caching strategies.

Abstract

Multi-Agent Path Finding (MAPF), which involves finding collision-free paths for multiple robots, is crucial in various applications. Lifelong MAPF, where targets are reassigned to agents as soon as they complete their initial targets, offers a more accurate approximation of real-world warehouse planning. In this paper, we present a novel mechanism named Caching-Augmented Lifelong MAPF (CAL-MAPF), designed to improve the performance of Lifelong MAPF. We have developed a new type of map grid called cache for temporary item storage and replacement, and created a locking mechanism to improve the planning solution's stability. A task assigner (TA) is designed for CAL-MAPF to allocate target locations to agents and control agent status in different situations. CAL-MAPF has been evaluated using various cache replacement policies and input task distributions. We have identified three main factors significantly impacting CAL-MAPF performance through experimentation: suitable input task distribution, high cache hit rate, and smooth traffic. In general, CAL-MAPF has demonstrated potential for performance improvements in certain task distributions, map and agent configurations.

Figures (6)

• Figure 1: Caching-Augmented Maps: (1) Blue grids represent Shelves $S$. (2) Purple grids represent Caches $C$. (3) Green grids represent unloading ports $U$. The upper map has multiple ports, while a single port is in the bottom one. In the multi-port map, each unloading port has an independent cache area, task queue, and agents. The cache areas are near the unloading ports within $\pm 2$ rows. For the single-port map, the port can utilize all agents and all caches. Given that the number of cache grids can affect the cache hit rate, we also tested different numbers of cache grids, ranging from 80 to 16, by removing cache grids column by column from right to left.
• Figure 2: Status 0: If the agent's task item is not in the cache when assigned, it must retrieve the item from a shelf. Status 1: If the agent's task item is in cache when assigned, it should retrieve the item from the cache. Status 2: If at least one writable cache exists when the agent retrieves the item from the shelf, the agent needs to insert the item into cache. Status 3: If no writable cache is available when an agent gets the item from the shelf, the agent goes to the unloading port. Status 4: After retrieving or inserting the item from/to the cache, the agent heads to the unloading port.
• Figure 3: The average frequency of agent wait actions on each map grid with 256 agents under Zhang distribution. As the agent number is large enough, both CAL-MAPF and baseline experience severe traffic congestion. The congestion position bias in the map could be caused by the randomization of item indexes and tasks in $Q$.
• Figure 4: Makespan (Bar chart, lower is better) and Cache Hit Rate (Line chart, higher is better). LRU, FIFO, and RANDOM represent CAL-MAPF with different cache replacement policies. NONE represents Lifelong MAPF without cache.
• Figure 5: Makespan (Bar chart, lower is better) and Cache Hit Rate (Line chart, higher is better). As the number of caches increases, CAL-MAPF's cache hit rate and makespan performance improve. In the MK, Real, and Zhang distributions, CAL-MAPF surpasses the baseline in most test settings. However, it is observable that as the number of agents increases, the improvement offered by CAL-MAPF diminishes. Additionally, an abnormal increase in makespan is observed at baseline under the Zhang distribution with 256 agents.