MECoBench
收藏资源简介:
MECoBench是由复旦大学、上海创新研究院及香港中文大学联合构建的多模态具身协作基准数据集,旨在系统评估多智能体在视觉接地环境中的合作能力。数据集包含192个测试案例,涵盖八类现实世界活动任务,支持并行与顺序两种协作结构,并集成多样化合作模式与通信机制,数据来源于VirtualHome家庭模拟器,通过场景接地与协作配置两阶段流程生成。该数据集主要应用于多模态大语言模型的具身智能研究,致力于解决多智能体在动态视觉观察下的协调、探索与通信等核心挑战,推动协作机器人在家庭辅助等实际场景中的应用。
MECoBench is a multimodal embodied collaboration benchmark dataset co-developed by Fudan University, Shanghai Institute of Innovation and The Chinese University of Hong Kong. It is designed to systematically evaluate the collaborative capabilities of multiple agents in visually grounded environments. The dataset includes 192 test cases covering eight categories of real-world activity tasks, supports two collaborative architectures: parallel and sequential, integrates diverse cooperation modes and communication mechanisms, and is generated via a two-stage workflow of scene grounding and collaboration configuration based on the VirtualHome household simulator. This dataset is primarily used for embodied intelligence research with multimodal large language models (LLMs), aiming to address core challenges such as coordination, exploration and communication of multiple agents under dynamic visual observations, and promote the application of collaborative robots in practical scenarios like home assistance.
MECoBench 数据集概述
项目名称:MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments
核心目标:系统性评估多智能体在具身环境中的协作能力。
平台与环境:
- 基于 VirtualHome 模拟环境。
- 使用经过修改的 VirtualHome v2.3.0 Unity 可执行文件(称为 MECoBench Simulator),支持多智能体房间限制和额外摄像头控制。
- 提供 Linux、macOS(Apple Silicon/Intel)、Windows(32/64位)等多种平台的模拟器版本。
任务数据集:
- 包含两种任务集合类型:
- 并行任务:
parallel.json - 顺序任务:
sequential.json
- 并行任务:
- 示例任务(小规模)位于
data/examples/目录:parallel_5.jsonsequential_5.json
- 完整任务集需从 Hugging Face 下载:
https://huggingface.co/datasets/q-i-n-g/MECoBench,置于data/task/目录下。
代码与资源:
- 代码仓库位于 GitHub:
https://github.com/q-i-n-g/MECoBench - 包含修改后的 VirtualHome Python API,支持多智能体交互。
- 移除内容:生成数据脚本、中间任务知识资产、本地笔记本、私有服务 URL 及机器特定启动脚本。
评估功能:
- 支持多种通信模式:
base(基础模式)no_comm(无通信)discuss_then_act(先讨论后行动)leader_worker(领导者-工作者模式)shared_memory(共享记忆)
- 可配置智能体数量(默认2个)、最大步数、模型别名(通过
eval/agents/models.json)。 - 输出结果:每个 case 的
result.json及批次级summary.json。
引用论文:
- arXiv:
https://arxiv.org/abs/2606.31966 - BibTeX 引用格式由仓库提供。




