Published by Crobotia, this dataset contains handheld bimanual demonstrations in real homes, covering garment unfolding, flattening and folding. Each episode includes left and right gripper videos, estimated camera poses, gripper opening angles, IMU, audio and multiple annotation layers.

Overview · Preview · Download & loading · Field reference

Dataset overview

Item	Content
Episodes	104,634
Duration	1,306.8 hours
Frames	141,134,078 (the synchronized pair counts once)
Size	Approximately 8.6 TB
Video	Left and right fisheye views, 960×960, 30 fps, H.264
Pose and gripper state	16-dimensional state / action per frame
IMU and audio	Approximately 1 kHz six-axis IMU and 48 kHz stereo audio per side
Annotations	Task intervals, action-step intervals and keyframe visual grounding
Calibration	Per-body camera intrinsics, camera–IMU extrinsics and IMU biases
Format	LeRobot v3.0, with an independent dataset per episode
License	CC BY-SA 4.0

Demonstration preview

The overlays illustrate estimated motion trajectories derived from the delivered poses. A white core indicates estimated jaw closure. Single-episode demonstration · Multi-scene gallery

Download and loading

This example downloads the first shard and reads one episode from it. Find other episodes and their shard and local paths in episodes.csv.

python -m pip install huggingface_hub 'lerobot==0.4.4' av
hf download Crobotia/iUMI README.md episodes.csv shards.csv checksums.sha256 \
  episodes/part-00000.tar --repo-type dataset --local-dir iUMI
tar -xf iUMI/episodes/part-00000.tar -C iUMI

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    repo_id="Crobotia/clothes_mixed_e84225",
    root="iUMI/data/clothes_mixed_e84225",
    video_backend="pyav",
)
item = ds[0]

See scripts/load_batch.py for reading multiple extracted episodes. IMU, audio and extra annotations are provided in each episode's corresponding files.

Package contents

Crobotia/iUMI
├── README.md / README_EN.md / README_zh.md / README_ZH.md
├── LICENSE / NOTICE / CHANGELOG.md
├── episodes.csv
├── shards.csv / checksums.sha256
├── assets/
├── qc/
├── scripts/load_batch.py
└── episodes/
    ├── part-00000.tar
    └── ...

Each tar shard contains complete, independent episodes. Extraction creates data/<episode>/, preserving the original meta/, data/, videos/, imu/, annotation/ and other files. Shards target approximately 3 GiB and follow whole-episode boundaries.

episodes.csv lists names, frames, durations, instructions, shard paths and extracted paths. shards.csv lists sizes, episode counts and SHA-256 hashes; checksums.sha256 covers the release files.

Fields and coordinate conventions

State: left side followed by right side, with eight values per side ordered as [x, y, z, qw, qx, qy, qz, gripper_deg]. Positions are in metres, quaternions are w-first and gripper angles are in degrees.
Action: the same layout as state, recording the next-frame state. The final row corresponds to the next frame in the source recording.
Time: timestamp = frame_index / 30, in seconds. Video frames and table rows correspond by frame index.
Pose: T_w_cam maps camera coordinates into that side's world frame. The camera-frame origin is the optical centre; x, y and z point right, down and forward respectively.
World frames: each side uses its own independent world frame, and episodes inherit the source recording's reference. Cross-arm relative geometry and precise camera–TCP extrinsics are outside this batch's calibration scope.
IMU: CSV files preserve raw angular velocity and standard specific force in the IMU body frame, in rad/s and m/s² respectively. Biases are provided as separate calibration parameters.
Gravity calibration: gravity_vector_camera_frame is the unit upward direction in the camera frame at the calibration reference orientation; gravity_magnitude_m_s2 is the gravity magnitude. R_cam_imu maps IMU-frame vectors into the camera frame. Fields are stored per side in meta/calibration.json.

Annotations and quality

Task intervals are human-written and reviewed. Action-step intervals and keyframe grounding are model-generated. Sources are recorded in meta/annotation.json; frame ranges follow [start, end).

All 104,634 episodes pass the publisher's inclusion checks, covering files, timing and media consistency, trajectories and annotations. Video content checks use periodic sampling. Results are available in qc/README.md.

Delivered poses are offline estimates, and model-generated annotations may contain boundary or localization inaccuracies. Images and audio retain the original recorded content, including collectors and home environments. Metadata is sanitized; faces and device identifiers are mosaicked in this page's display media.

License and citation

Released under CC BY-SA 4.0. The full license is in LICENSE, attribution information in NOTICE, and revisions in CHANGELOG.md. The attribution name is Crobotia.

@dataset{crobotia_iumi_2026,
  title  = {Crobotia Household Bimanual Manipulation Challenge Dataset (iUMI)},
  author = {{Crobotia}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Crobotia/iUMI},
  note   = {104634 episodes; 1306.8 hours}
}

Downloads last month: 24

Total file size:

7.73 TB