PRISM
收藏资源简介:
PRISM(Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing)是由北京大学等机构联合创建的大规模多模态数据集,专注于高精度、接触密集的工业操作场景。该数据集包含超过5000条机器人轨迹(总计45小时遥操作演示),同步采集多视角RGB-D、六轴力/力矩、触觉及机器人状态等模态数据,涉及25种以上操作任务(如电子元件插拔、传送带分拣等)。数据通过外骨骼、追踪器和VR三种遥操作方式收集,覆盖多种机器人平台和末端执行器,并引入人为扰动以增强鲁棒性。PRISM旨在为学习接触密集、长时程的工业操作提供高保真基准,推动多模态感知与控制策略在真实制造环境中的泛化应用。
PRISM (Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing) is a large-scale multimodal dataset jointly developed by Peking University and other research institutions, focusing on high-precision, contact-intensive industrial manipulation scenarios. This dataset contains over 5,000 robot trajectories, with a total of 45 hours of teleoperation demonstrations, and synchronously acquires multimodal data including multi-view RGB-D, six-axis force/torque, tactile sensing, and robot state information. It covers more than 25 types of manipulation tasks, such as electronic component insertion and extraction, conveyor belt sorting, and other typical industrial operations. The data is collected via three teleoperation modalities: exoskeleton, tracking device, and VR, and covers diverse robot platforms and end-effectors. Human-induced perturbations are incorporated into the dataset to enhance its robustness. PRISM aims to provide a high-fidelity benchmark for learning contact-intensive, long-duration industrial manipulation tasks, and to facilitate the generalized deployment of multimodal perception and control strategies in real-world manufacturing environments.
PRISM 数据集详情总结
数据集概述
PRISM(Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing)是一个面向接触丰富型真实工业操作的大规模多模态数据集,由北京大学等多家机构联合发布。该数据集旨在弥补现有机器人学习数据集在短视域、低接触任务(如抓取放置)方面的不足,重点关注工业装配所需的精密度控制、力/力矩或触觉调节以及多模态反馈。
数据规模
- 工业任务:25+ 项,涵盖电子元件插拔、传送带分拣等操作
- 机器人轨迹:5,000+ 条,并配有对应的人类示范
- 数据总时长:45+ 小时
- 图像数据:约 2,700 万张
- 感知模态:多视角 RGB-D 视觉、6自由度力/力矩传感、触觉传感、本体感知
数据采集平台
数据集通过三种互补的遥操作界面采集数据:
- 外骨骼控制
- 追踪器控制
- VR 控制
这些平台引入了不同的人类控制风格,降低了单一采集设置的偏差。每个轨迹均记录了机器人状态、多视角 RGB-D 图像、力/力矩测量、触觉观测、夹爪状态及时间戳,以实现多模态对齐。
数据特点
- 覆盖多机器人形态、多视角观测、多模态传感流及多样化工业技能
- 每个轨迹同步记录视觉、触觉、力/力矩和本体感知信号
- 专门针对高精度工业操作中的接触、力调节和多模态反馈需求设计
- 适用于多模态感知与控制研究,为接触丰富型、可泛化的真实制造环境操作提供基础
发布信息
- 论文已提交至 arXiv(编号:2608.17962)
- 数据集即将发布
引用格式: bibtex @misc{yu2026prismprecisioncontactrichrealworld, title={PRISM: Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing}, author={Tengbo Yu and Jiahao Wu and Hanning Wang and Rui Chen and Chuanhou Liu and Chuang Sun and Hangxin Liu}, year={2026}, eprint={2608.17962}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2608.17962}, }

- 1PRISM: Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing北京大学·通用人工智能国家重点实验室; 台达智能; 北京大学武汉人工智能研究院; 湖北人形机器人创新中心有限公司; 中国信息通信研究院 · 2026年



