ORS3D-60K
收藏资源简介:
ORS3D-60K是由华中科技大学与小米公司联合构建的大规模具身智能数据集,专注于融合运筹学知识的3D任务调度。该数据集包含60,825项复合任务,覆盖4,376个真实室内场景,每个任务平均包含311个词汇的解决方案描述。数据源自ScanNet、HM3D等五大真实3D场景数据库,通过GPT-4o生成任务元信息并经过优化求解器计算最优调度方案。该数据集旨在解决具身智能在三维物理世界中并行执行多任务的效率优化问题,推动语言理解、空间感知与运筹决策的协同发展。
ORS3D-60K is a large-scale embodied intelligence dataset jointly constructed by Huazhong University of Science and Technology and Xiaomi Corporation, specializing in 3D task scheduling integrated with operations research knowledge. It contains 60,825 complex tasks covering 4,376 real indoor scenes, with the average solution description of each task comprising 311 vocabulary terms. The dataset is sourced from five real 3D scene databases including ScanNet and HM3D, where task metadata was generated via GPT-4o and optimal scheduling schemes were calculated using optimization solvers. This dataset aims to address the efficiency optimization issue of embodied intelligence when executing multiple parallel tasks in the 3D physical world, and advance the synergistic development of language understanding, spatial perception and operational decision-making.
数据集概述
基本信息
- 数据集名称: ORS3D-60K
- 任务类型: 基于运筹学知识的3D基础任务调度
- 数据规模: 60,000个复合任务,覆盖4,000个真实世界场景
- 发布机构: 华中科技大学、小米MiLM Plus
- 相关论文: Cook and Clean Together: Teaching Embodied Agents for Parallel Task Execution
任务特点
- 要求智能体通过语言理解、3D空间基础和效率优化的协同作用
- 强调利用可并行子任务最小化总完成时间
- 支持在3D物理世界中执行自然语言指令和高效动作
数据内容
- 包含大规模复合任务数据集
- 涵盖多种真实世界3D场景
- 支持语言理解、3D基础和调度效率的多模态评估
获取方式
- 数据集下载地址: https://huggingface.co/datasets/H-EmbodVis/ORS3D-60K
- 3D场景数据来源: https://github.com/scene-verse/SceneVerse/blob/main/DATA.md
发布时间
- 代码和数据集发布时间: 2025年11月24日
- 相关论文被AAAI 2026接收为口头报告(接收率约4.5%)




