dexbench/rlds

DexBench是一个灵巧操作演示数据集，通过Isaac Lab重放，打包为Open-X-Embodiment风格的RLDS/TFDS数据集。每个帧包含：256×256分辨率的第三人称和手腕RGB图像、本体感知、关节空间动作以及自然语言指令。数据集包含两个变体：单手机器人变体（735个episodes，15个任务，动作/状态维度28，手腕相机为wrist_image）和双手机器人变体（746个episodes，9个任务，动作/状态维度56，手腕相机包括left_wrist_image和right_wrist_image）。总计：1,481个episodes，约585,000帧，30 fps，256×256 RGB。每个episode的视觉随机化被禁用，即HDRI背景和桌面纹理在所有episodes中固定，以确保演示的视觉稳定性。数据集适用于机器人学习和强化学习研究。DexBench dexterous manipulation demonstrations replayed through Isaac Lab, packaged as Open-X-Embodiment-style RLDS / TFDS datasets. Each frame includes: third-person and wrist RGB images at 256×256 resolution, proprioception, joint-space action, and a natural-language instruction. The dataset contains two variants: a single-hand variant (735 episodes, 15 tasks, action/state dimension 28, wrist camera as wrist_image) and a bimanual variant (746 episodes, 9 tasks, action/state dimension 56, wrist cameras as left_wrist_image and right_wrist_image). Total: 1,481 episodes, approximately 585k frames, 30 fps, 256×256 RGB. Per-episode visual randomization is disabled, meaning the HDRI background and table texture are fixed across all episodes for visually stable demonstrations. The dataset is suitable for robotics and reinforcement learning research.