ZenO App Household
7 first-person recordings of ordinary household work, with both hands tracked in three dimensions and in real metres, laid out in LeRobot v2.1.
| Episodes | 7 |
| Frames | 101,205 at 30 fps, about 56.2 minutes |
| Video | observation.images.head, 1920x1440 |
| State | 7 floats, camera position and orientation |
| Action | 20 floats, both wrists and both grippers |
| Coordinate frame | ROS REP 103, X forward, Y left, Z up, metric |
Watch it first
Open in the LeRobot dataset visualiser
Scrub any episode in the browser, with the wrist and gripper channels plotted under the video. Nothing to install.
Why the depth matters
These were recorded on an iPhone with a LiDAR sensor. Depth is measured rather than inferred, so hand positions are in real metres in a gravity aligned world frame, not in pixels and not up to an unknown scale. A trajectory can be compared between two recordings, or handed to a robot, without first solving for how big the scene was.
Hand keypoints come from MediaPipe Hands. Each keypoint is back projected through the LiDAR depth map and the camera pose that ARKit reports for that frame, so what you get is a position in the room rather than a position in the picture.
The action vector
Twenty numbers per frame, ten for each hand, in this order:
| Index | Channel | Meaning |
|---|---|---|
| 0 to 2 | l_pos_x/y/z |
left wrist position, metres |
| 3 to 8 | l_rot6d_0..5 |
left wrist orientation, first two columns of the rotation matrix |
| 9 | l_gripper |
left hand aperture, 0 closed to 1 open |
| 10 to 12 | r_pos_x/y/z |
right wrist position, metres |
| 13 to 18 | r_rot6d_0..5 |
right wrist orientation |
| 19 | r_gripper |
right hand aperture |
Rotation is carried as a 6D representation rather than a quaternion because it is continuous, which matters when a network has to regress it.
action.left_valid and action.right_valid are booleans. A hand that was not
detected in a frame carries its last known pose, so check the flag rather than
looking for a sentinel value.
About the gripper channel
It is hand aperture, the thumb to index distance divided by hand span, mapped
through one fixed pair of bounds for every episode: 0.12 reads as closed and
0.66 as open. Those bounds are the 5th and 95th percentile of aperture measured
across 25,011 detections and they are recorded in meta/info.json.
It is deliberately not a grasp label. Aperture says how open the hand is, not whether it closed on anything, and the two disagree often enough that treating one as the other would be wrong. If you need contact events, derive them and validate them yourself.
Quality
Every episode here passed four automatic checks before it was included:
| Check | What it asks |
|---|---|
| Revisit drift | When the camera returns to a place it has been, does that place land where it did before |
| Bone length | Do the finger segment lengths stay constant frame to frame |
| Speed | Do the hands and camera move at speeds a person can produce |
| Inertial agreement | Does the phone's own motion sensor agree with the pose track, and with what time offset |
Recordings that fail any of them are held back rather than deleted. The checks are conservative and a failure usually means the tracking slipped, not that the recording is worthless.
1 recording(s) were excluded from this build by those checks.
Quick start
pip install lerobot
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("zeno-labs/zeno-app-household")
sample = ds[0]
sample["observation.state"] # [x, y, z, qx, qy, qz, qw]
sample["action"] # 20 floats, see the table above
sample["observation.images.head"] # [C, H, W] uint8
What this is not
- There is no robot in it. The action is where a person's hands went, not a joint command. Retargeting to a specific arm is left to you.
- Body pose is not included. Hands and head only.
- Hand detection is not perfect. Detection rates per episode run from 94 to 99 percent, and hands drop out when they leave frame, go behind an object, or go into water.
- Episodes may be republished. Pin a commit if you need strict reproducibility.
Consent and privacy
Every recording was contributed knowingly by a person who was paid for it, under terms that cover research and commercial use of the resulting data. Faces and other identifying detail are removed before a recording enters the pipeline.
Related
Same capture setup and the same schema, different recordings: zeno-labs/egostation-iphone-lidar-household-v1. Neither replaces the other.
Licence
CC BY-NC 4.0, non commercial. For the full corpus or commercial terms, contact support@zen-o.xyz.
- Downloads last month
- 60