cloudwalk-research/psi0-g1-he-pickplace-9x40ep

该数据集是一个用于人形机器人（Unitree G1）抓取和放置任务的数据集，基于Humanoid-Everyday（HE）数据集重新映射而来。原始数据使用Dex3-1三指手（每手7自由度），但本数据集将其转换为适用于Inspire RH56DFTP六指手的锁步模式（每手一个开/关标量广播到6个通道），以便与LeRobot v2.1框架和Psi0等策略直接兼容。数据集包含9个不同的单手机器人任务（4个右手主动，5个左手主动），每个任务有40个片段，总计360个片段和121,875帧图像（30fps）。数据格式包括自我中心视角的RGB视频（480x640）以及动作（36维）和状态（32维）向量，动作维度涵盖手部、手臂、躯干和移动控制。数据集中详细说明了手部动作的转换方法（如基于百分位数的归一化、滞后二值化处理），并修复了早期版本中的右手极性反转问题。该数据集主要用于视觉-语言-动作（VLA）和模仿学习研究，特别是针对人形机器人的抓取操作任务。This dataset is for humanoid robot (Unitree G1) pick-and-place tasks, remapped from the Humanoid-Everyday (HE) dataset. The original data uses Dex3-1 three-finger hands (7 DoF per hand), but this dataset converts it to a lockstep schema compatible with Inspire RH56DFTP six-finger hands (one open/close scalar broadcast to 6 channels per hand), enabling direct integration with the LeRobot v2.1 framework and policies like Psi0. It includes 9 different unimanual tasks (4 RH-active, 5 LH-active), with 40 episodes per task, totaling 360 episodes and 121,875 frames (30fps). The data format includes egocentric RGB videos (480x640), action vectors (36-D), and state vectors (32-D), covering hand, arm, torso, and locomotion controls. The dataset details hand-action conversion methods (e.g., percentile-based normalization, hysteresis binarization) and fixes right-hand polarity inversion issues from earlier versions. It is designed for vision-language-action (VLA) and imitation-learning research, specifically for humanoid manipulation tasks.