# 50000-Hours-UMI-Embodied-Manipulation-Dataset ## Overview UMI Embodied Manipulation Dataset: a large-scale real-world teleoperation dataset for embodied AI, captured in real homes with uDAS 1.0. It offers ≈50,000 h footage, ≈934k episodes and ≈19.2M action segments. Each episode records six first-person streams (head 1408×1280, wrist 640×480, 30 fps) plus 30 Hz aligned proprioceptive state. Two-level annotations, scene/object/action labels, LeRobot v2.1. For more details, please refer to the link: https://www.nexdata.ai/datasets/embodied-ai/2300?source=Huggingface ## Specifications ### Application domain Embodied AI, robot manipulation, visionlanguage-action models ### Task type Everyday household tasks (cleaning, tidying, organizing, cooking, etc.) ### Capture method Human operation, headset localization + handheld grippers, six synchronized cameras ### Capture environment Real homes, not a laboratory ### Observation modalities 6 RGB video streams + bimanual end-effector poses, gripper opening, torso pose ### Annotation coverage High-level task description + subtask timeline ### Supported algorithms VLA pre-training and fine-tuning, imitation learning, action tokenization, long-horizon task planning, generalization evaluation ## Licensing Information Commercial License