OmniSharing DB 帕西尼全模态具身智能数据集

OmniSharing DB 帕西尼全模态具身智能数据集，是国内首个融合多维触觉、视觉、关节角度、动作轨迹、语音、文本等全维度人类物理交互行为的具身智能数据集，可为具身智能体提供更丰富、更全面的环境信息，使其能够更精准地理解任务、感知世界。该数据集采用帕西尼“人因”全模态数据采集系统，通过集成自研 PMEC 触觉数据手套、空间多维视觉阵列和语音采集模块，基于真人自然动作姿态，高效、精准地实现力触觉、摩擦、形变等物理交互行为的全方位捕获。并通过 Neural Mesh 多模态神经织网技术与 SomaRedirect 体感重定向系统，其可实现全模态数据的无损采集与跨本体泛化能力。这一数据生产方式不仅无需依赖昂贵的机器人本体进行数据采集，具备高效率、低成本的显著优势，同时使用这些数据训练出的具身智能大模型，在面对不同构型的机器人时，依旧能展现出卓越的适应性和执行力。该数据集预计年产近 2 亿条高质量、全模态、标准化、高可用性的具身智能数据，不仅解决了行业长期面临的具身智能“数据质量与生产效率”的双重困境，更为推动建设国家级数据要素平台、强化数据基础设施提供了有力支持。

OmniSharing DB: Pacini All-Modal Embodied Intelligence Dataset is the first domestically developed embodied intelligence dataset in China that integrates full-dimensional human physical interaction behaviors including multi-dimensional tactile sensation, vision, joint angles, motion trajectories, speech and text. It can provide embodied AI agents with richer and more comprehensive environmental information, enabling them to understand tasks and perceive the world more accurately.
This dataset adopts the Pacini "human-centric" all-modal data acquisition system. By integrating the self-developed PMEC tactile data glove, spatial multi-dimensional vision array and speech acquisition module, based on natural human motion postures, it efficiently and accurately captures all-round physical interaction behaviors such as force-tactile sensation, friction and deformation. Additionally, through the Neural Mesh multi-modal neural web technology and SomaRedirect somatosensory redirection system, it enables lossless acquisition of all-modal data and cross-embodiment generalization capability.
This data production method does not rely on expensive robot platforms for data collection, and has significant advantages of high efficiency and low cost. Meanwhile, embodied intelligence LLMs trained on this data can still exhibit excellent adaptability and execution performance when facing robots with different configurations.
This dataset is expected to produce nearly 200 million high-quality, all-modal, standardized and highly available embodied intelligence data entries per year. It not only solves the dual dilemma of "data quality and production efficiency" in embodied intelligence that the industry has long faced, but also provides strong support for promoting the construction of national-level data element platforms and strengthening data infrastructure.

搜集汇总

背景与挑战

背景概述

OmniSharing DB是国内首个全模态具身智能数据集，融合触觉、视觉、动作等多维交互数据，采用自研采集系统实现高效低成本生产，年产2亿条高质量数据，解决行业数据质量与生产效率的双重困境。