RoboMIND机器人数据集

本数据集面向机器人操作大模型研发团队、具身智能算法研究机构及机器人技能学习系统开发商，旨在解决异构机器人数据采集中构型覆盖不全、任务多样性不足及仿真迁移困难等核心问题。数据集总含55,000条机器人轨迹数据，涵盖Franka Emika Panda单臂、“天工”人形机器人、AgileX Cobot Magic V2.0双臂及UR-5e单臂等多种构型，末端执行器包括夹爪与灵巧手。覆盖279项任务、61种物体，涉及家居、厨房、工厂、办公、零售等主要生活服务场景，并包含根据真实场景搭建的数字孪生数据。与传统脚本采集不同，本数据集采用自研遥操作系统与动作捕捉技术，结合云原生平台与三级质检体系，确保了数据的高真实度与低误差。

This dataset is targeted at R&D teams developing large language models for robotic manipulation, embodied intelligence algorithm research institutions, and robot skill learning system developers, aiming to address core issues in heterogeneous robot data collection, such as incomplete coverage of robot configurations, insufficient task diversity, and difficulties in simulation-to-real transfer. The dataset contains a total of 55,000 robot trajectory entries, covering various robot configurations including the Franka Emika Panda single-arm robot, the "Tiangong" humanoid robot, the AgileX Cobot Magic V2.0 dual-arm robot, and the UR-5e single-arm robot, with end-effectors including grippers and dexterous hands. It covers 279 tasks and 61 types of objects, involving major daily service scenarios such as home, kitchen, factory, office, and retail, and includes digital twin data built based on real-world scenes. Unlike traditional script-based data collection, this dataset adopts a self-developed teleoperation system and motion capture technology, combined with a cloud-native platform and a three-level quality inspection system, to ensure high authenticity and low error rate of the data.

搜集汇总

背景与挑战

背景概述

RoboMIND机器人数据集是一个面向机器人操作大模型与具身智能研究的多模态数据集，包含55,000条机器人轨迹数据，覆盖多种机器人构型、279项任务和61种物体，涉及家居、厨房等主要生活服务场景。它采用自研遥操作与动作捕捉技术采集，旨在解决异构机器人数据采集中构型覆盖不全、任务多样性不足及仿真迁移困难等问题，为机器人泛化能力研究提供关键支撑。