Enhancing Lifelong Multi-Agent Path Finding with Cache Mechanism

TL;DR

The paper tackles Lifelong MAPF in warehouse environments by introducing L-MAPF-CM, which integrates a cache-based intermediate storage with low-level path planning. A Task Assigner (TA) collaborates with a cache lock mechanism to manage dynamic target allocation and safe concurrent cache access. Empirical results show that higher cache hit rates and traffic smoothness yield notable throughput gains, though benefits taper as agent counts rise and congestion emerges in dense scenarios. The work highlights practical avenues for improving warehouse routing, including smarter task ordering, data-driven cache policies, and hierarchical caching structures.

Abstract

Multi-Agent Path Finding (MAPF), which focuses on finding collision-free paths for multiple robots, is crucial in autonomous warehouse operations. Lifelong MAPF (L-MAPF), where agents are continuously reassigned new targets upon completing their current tasks, offers a more realistic approximation of real-world warehouse scenarios. While cache storage systems can enhance efficiency and reduce operational costs, existing approaches primarily rely on expectations and mathematical models, often without adequately addressing the challenges of multi-robot planning and execution. In this paper, we introduce a novel mechanism called Lifelong MAPF with Cache Mechanism (L-MAPF-CM), which integrates high-level cache storage with low-level path planning. We have involved a new type of map grid called cache for temporary item storage. Additionally, we involved a task assigner (TA) with a locking mechanism to bridge the gap between the new cache grid and L-MAPF algorithm. The TA dynamically allocates target locations to agents based on their status in various scenarios. We evaluated L-MAPF-CM using different cache replacement policies and task distributions. L-MAPF-CM has demonstrated performance improvements particularly with high cache hit rates and smooth traffic conditions.

Figures (7)

• Figure 1: Warehouse 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 is associated with its own independent cache area, task queue, and set of agents. The cache areas are located near the unloading ports, within a range of $\pm 2$ rows. In contrast, the single-port map allows the port to utilize all available agents and caches. Given that the number of cache grids can influence the cache hit rate, we tested different configurations by varying the number of cache grids from 80 to 16. This was achieved by progressively removing cache grids column by column from right to left.
• Figure 2: State Machine: (a) SF_GET: The agent is moving to a shelf to retrieve its task item. (b) CA_MOV: The agent is moving to a cache to remove all items. (c) CA_GET: The agent is moving to a cache to retrieve a task item. (d) CA_ADD: The agent is moving to a cache to store task items. (e) SF_ADD: The agent is moving to a shelf to return items from a cache. (f) UP_END: The agent is moving to its unloading port with one task item.
• Figure 3: Status examples: The blue circle represents agents, the purple square represents items to be placed, and the red square represents items to be taken out.
• Figure 4: Throughput (Bar chart, higher is better) and Cache Hit Rate (Line chart, higher is better). LRU, FIFO, and RANDOM represent L-MAPF-CM with different cache replacement policies. NONE represents L-MAPF algorithm without cache.
• Figure 5: The frequency of agent wait actions on each map grid with 256 agents and 80 caches under Zhang distribution: Given the large number of agents, both L-MAPF-CM and the baseline experience significant traffic congestion (dark color area). The bias in congestion positions across the map is attributed to item indices and tasks in $Q$.