Sample clip: three drones to a humanoid.

Eight seconds of a jog with a 90° turn, captured by three off-the-shelf drones with no markers, timecode or calibration target. Every file below is what a customer receives per clip. Free under CC BY 4.0.

Download the files Explore it 3D hands sample

Drone 1Drone 2Drone 3Foot contact

3D reconstruction · metres · z up

Drone 1

Drone 2

Drone 3

0.00 s

Drag the timeline. Orange skeletons on the videos are the 3D reconstruction projected back into each drone. Lanes: heel and ball contact for each foot, then held-out error.

59 / 59Randomized runs completed: pushes, friction, mass, motor gains, 0–20 ms latency

2.9 cmRobot keypoint error against the captured motion

0.10 mMean path error over the clip

–Held-out-camera error, median across drones and joints

Recovered from the footage

What the pipeline solved

–Drone 2 clock offset, recovered from the footage

–Drone 3 clock offset, recovered from the footage

3 × 240Moving camera poses, one per drone per frame

1.83 mMetric scale from subject height; ground at z = 0

Download · CC BY 4.0

Clip contents

File	Contents	Size
drone1.mp4	Drone 1 video, 1920 × 1080, 29.97 fps	11.2 MB
drone2.mp4	Drone 2 video, 1920 × 1080, 29.97 fps	13.9 MB
drone3.mp4	Drone 3 video, 1920 × 1080, 29.97 fps	16.3 MB
calibration/cameras.json	Intrinsics, radial k1, recovered clock offsets, keyframe flags per drone	9 KB
calibration/extrinsics.npz	Per-frame R, t for each moving drone (3 × 240)	33 KB
motion/body_coco17.npz	COCO-17 joints in metres (240 × 17 × 3) and each drone’s 2D detections	123 KB
motion/body_fit_joints.npz	Parametric body fit joints, smoothed	304 KB
motion/contacts.npz	Per-frame planted flags: left/right heel and ball	2 KB
robot/g1_qpos.npz	Unitree G1 retargeted trajectory, 29 DoF + floating base	66 KB
robot/g1_motion_50hz.csv	G1 trajectory resampled to 50 Hz	136 KB
robot/physics_eval.json	MuJoCo evaluation: 60 runs, randomization ranges, fall rule, errors	69 KB
qc/error_report.json	Held-out-camera error: per drone, per joint, percentiles	1 KB
qc/heldout_residuals.npz	Per-frame held-out residuals in px and mm (3 × 240 × 17)	88 KB
metadata.json	Capture metadata and licence	1 KB
LICENSE.txt	Licences: CC BY 4.0 for this clip, evaluation licence for the 3D hands sample	1 KB

World frame for every file: metres, z up, ground at z = 0, right-handed. Cameras: x_cam = R x_world + t, OpenCV axes, intrinsics at 1920 × 1080. Attribution: “Autelica (autelica.com)”. The motion library and pilot data are licensed separately.

Sample data · evaluation licence

3D hands sample

Five desk tasks filmed by a head camera and three fixed cameras: both hands as 21 joints in one world frame, with the head camera’s pose on every frame.

Download all (zip, 21.1 MB)Read the hands article

Box stackingStacks small product boxes into piles, then packs them into a bag on the floor.15.0 s · hands 46 % left, 52 % right · held-out 20.9 mm MP4 4.3 MB JSON 271 KB NPZ 180 KB

Laptop: open, type, closeOpens a laptop, types briefly, closes it and moves it aside.15.0 s · hands 63 % left, 92 % right · held-out 10.8 mm MP4 3.6 MB JSON 372 KB NPZ 227 KB

USB-C into a laptopPlugs a USB-C cable into the side ports of a closed laptop and unplugs it, twice.16.0 s · hands 68 % left, 97 % right · held-out 12.2 mm MP4 4.0 MB JSON 415 KB NPZ 250 KB

Mount and thumbscrewTakes a mount and a thumbscrew from a case, screws them together with both hands, packs them away and zips the case.16.0 s · hands 75 % left, 88 % right · held-out 13.7 mm MP4 3.5 MB JSON 416 KB NPZ 264 KB

Battery packs out and inPulls the battery packs out of three small drones and pushes them back in.16.0 s · hands 72 % left, 89 % right · held-out 13.2 mm MP4 3.9 MB JSON 405 KB NPZ 263 KB

Joints: 21 per hand: the wrist, then 4 per finger from base to tip. Left and right drawn in different colours.
Frame and units: One world frame for all clips: origin on a tag on the desk, z up, millimetres. Camera poses as 4 × 4 world-from-camera matrices.
Frame rate: 30 fps, 2340 frames in total, 78 s. Frame i of the data is frame i of the video.
Cameras: One head-mounted and three fixed consumer action cameras, fisheye intrinsics for each. No markers on the hands.
Head tracking: Head camera pose on 98–100 % of frames, shown as a moving frustum and path in each video.
Per frame: Joint positions, a confidence tier and a visibility flag per hand, head camera pose, timestamps.
Error report: Held-out camera check per clip: each fixed camera left out in turn and compared with the hands measured from the others. Median 10.8–20.9 mm.
Files: MP4 (H.264), JSON and NPZ per clip; README with the full spec; licence.

Clips chosen by one rule: one clip per task type, and within a type the clip with the fastest head motion (90th percentile of the head camera's rotation speed); the clip with the fastest head motion of the session is always included. Pouring from a bottle is left out for privacy. The capture is described in Eight tabletop tasks seen by four cameras. Full datasets and other licences: hello@autelica.com.

For frontier labs

How teams use clips like this

Pretraining data

Measured whole-body motion, not single-camera estimates
Fast, outdoor and work-site motion headcams miss
Per-clip quality grade for data weighting

Evaluation sets

Held-out sites and tasks with metric ground truth
Physics validation on your embodiment
Repeatable protocol and seeds

Label QA for egocentric data

Exocentric cameras alongside head cameras
Metric reference for auto-labelled body and hands
Measured label error per session

Site-specific deployment

Capture at the facility before go-live
Motion and evaluation per site
Repeat capture per rollout

Request a pilot Full specifications