Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Free to download. We ask six short questions because this is an early release and your answers decide what we build next — resolution, labels, and how much data we collect. Access is granted automatically on submit.
Log in or Sign Up to review the conditions and access this dataset content.
Egocentric Wrist-View Camera + IMU Data for Physical AI
Wrist-mounted camera + synchronized 6-axis IMU. The view a robot's wrist camera actually has — close to the hand, moving with it, object filling the frame.
Free 30-episode sample. Task: folding and organizing cloth. LeRobot v3.0, loads in one line.
Why wrist view
Most egocentric datasets are head-mounted: they watch the hand from the outside, several tens of centimeters away, with the whole room in frame. A robot arm's wrist camera does not see that. It sees the object it is about to touch, filling the frame, moving as the hand moves.
That gap is why head-mounted data transfers poorly to wrist-camera policies. This data is recorded from the wrist.
What it's for
- Pre-training visual encoders for wrist-camera manipulation policies
- Studying hand–object interaction from the manipulator's own viewpoint
- Sim-to-real work that needs the real optical and motion characteristics of a wrist-mounted sensor (rolling shutter, motion blur, fixed focus)
What's inside
| Episodes | 30 (sample) · 300 in the full pack |
| Length | ~19.9 s each (598 frames @ 30 fps) |
| Video | 640×480, 30 fps CFR |
| IMU | LSM6DSO, 6-axis, 104 Hz raw |
| Format | LeRobot v3.0 |
What this is NOT
Read this before downloading.
- No action labels. These are human demonstrations, not robot teleoperation. If your pipeline needs actions, you must derive them (hand pose estimation → retargeting). We have not done this. See What you can derive without labels below for what the video and IMU give you on their own.
- No gripper states, object poses, or language annotations.
- No audio track.
What you can derive without labels
No action labels does not mean the video is inert. The camera and the IMU sit on the same rigid body, so wrist motion is measured rather than inferred from the image.
The figure above was produced from a single episode in this repository using nothing but the files it ships with: dense optical flow, subtraction of the median flow field, and a texture mask to suppress low-contrast regions where flow is unreliable. What the residual measures is how far the view departs from moving as one piece. The lower panel puts it beside wrist motion from the IMU gyro; the shaded bands are where that departure is largest.
Read it as the difference between carrying and working, not as contact. Grip an object and swing the arm, and everything in view travels together, so the residual falls even though contact never breaks. Press the object, fold it, or turn it over, and the residual rises. On one clip we measured global flow outside the shaded bands at 90% of its value inside them — the hand never let go, it was simply carrying rather than working.
One caveat worth stating plainly: a handled object often covers more than half the frame, and the median flow then follows the object rather than the camera. The "departure from the dominant motion" reading holds either way, which is why we state it that way rather than claiming the camera's motion has been removed.
This decomposition is well-posed here in a way it is not for head-mounted footage. There, camera motion has to be recovered from the image itself — which is exactly what breaks down when a hand and a deforming object fill the frame.
The script is about 200 lines of OpenCV. We are happy to share it — just ask.
Technical notes
- This sample is a curated selection from the full pack.
- Fixed-focus optics — no autofocus hardware, no focus hunting mid-clip.
- Source video is variable frame rate (~26.4 fps), resampled to 30 fps CFR. Frames are duplicated, never dropped; timing error stays within ±17 ms.
- Video/IMU sync offset was measured by cross-correlation (+548 ms) and applied.
observation.stateis filtered for single-sample sensor read artifacts.imu_raw/keeps the unfiltered 104 Hz values.
How the sample was chosen
Selection is measured, not eyeballed. Each clip is scored and the sample is spread across capture devices:
| weight | measure | |
|---|---|---|
| Sharpness | 55% | variance of Laplacian |
| Exposure | 30% | mean luma near mid-grey, few clipped pixels |
| Motion | 15% | gyro σ — mid-range scores highest; too fast means motion blur |
Spreading across devices matters more than raw score — a sample from a single wearer would not tell you much about the pack.
Load it
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("inhandplus/egocentric-wrist-view-camera-IMU-data")
print(ds.meta.info)
print(ds[0]["observation.state"]) # 6-axis IMU at frame time
imu_raw/ holds the untouched 104 Hz CSVs if you need the original rate.
Collection and consent
Recorded by consenting participants in home environments on our own hardware. Clips were reviewed one by one before release; anything with identifiable people, screens, or documents was removed.
Full pack
300 episodes with a commercial, non-exclusive license: https://inhandplus.com/data/task/cloth
Custom collection for a specific manipulation is also available — we can record a new task in about a week.
Citation
@misc{inhandplus2026wristcloth,
title = {Wrist Cloth Manipulation: first-person wrist camera with synchronized IMU},
author = {InHandPlus},
year = {2026},
url = {https://huggingface.co/datasets/inhandplus/egocentric-wrist-view-camera-IMU-data}
}
Contact
biz@inhandplus.com — questions go to the people who built the pipeline.
- Downloads last month
- 55