HyPA-MAPF-Benchmark

Dataset supporting the paper “Validation-Driven Assessment of Large Language Models and HyPA-MAPF for Multi-Agent Path Finding in Constrained Industrial Scenarios”.

This dataset contains the benchmark instances, prompts, model outputs, validation results, selected solver information, makespan values, and generated paths used in the experimental evaluation. The benchmark includes 100 grid-based Multi-Agent Path Finding scenarios under the disappear-at-goal convention, divided into five categories: simple no-conflict, open-space crossing, narrow corridor crossing, side-pocket required, and three-robot bottleneck.

The dataset includes results for Independent A-star, Sequential Priority A-star, the proposed HyPA-MAPF architecture, and three Large Language Models: Gemini, Claude, and ChatGPT. For each scenario, the dataset provides the prompt given to the LLMs, the raw generated paths, the validated solution paths for the algorithmic methods, success or failure indicators, makespan values, and the selected HyPA-MAPF solver.

The material is intended to support reproducibility, independent validation, and further research on validation-driven assessment of LLMs and hybrid planning methods for Multi-Agent Path Finding.

本数据集支撑论文《受限工业场景下多智能体路径规划（Multi-Agent Path Finding，MAPF）的大语言模型与HyPA-MAPF架构验证驱动评估》。本数据集包含实验评估所用的基准测试用例、提示词（Prompt）、模型输出结果、验证结果、选定求解器信息、完工时间（makespan）值以及生成的路径。该基准数据集包含100个基于栅格的多智能体路径规划场景，遵循“到达目标后消失”的规则，分为五大类别：无冲突简单场景、开放空间穿越场景、狭窄通道穿越场景、侧袋式需求场景以及三智能体瓶颈场景。本数据集涵盖了独立A*算法（Independent A-star）、序列优先级A*算法（Sequential Priority A-star）、本文提出的HyPA-MAPF架构以及三款大语言模型（Large Language Model，LLM）：Gemini、Claude、ChatGPT的实验结果。针对每个场景，本数据集均提供了向大语言模型输入的提示词、原始生成路径、各类算法方法的验证后求解路径、成功/失败判定指标、完工时间值以及选定的HyPA-MAPF求解器信息。本数据集旨在支撑多智能体路径规划领域中，面向大语言模型与混合规划方法的验证驱动评估相关研究的可复现性、独立验证以及后续拓展研究。