A capture lab for robot hand data.
We record people doing real manual work with a calibrated RGB-D sensor, lift both hands to metric 3D, and report the error in millimeters. Every episode is reviewed by a person before it leaves the lab.
How an episode moves through the lab
The same five stations run in ACTEM Studio, our internal capture software. Each one leaves a file you can check.
- 1
Calibrate
Serial, firmware and factory intrinsics are logged before each session.
calibration.json - 2
Capture
Clean RGB, raw depth and IMU with device timestamps, written straight to disk.
rgb.mp4 · depth.zarr - 3
Lift to 3D
Hand pose on clean RGB, lifted with sensor depth. Never inferred depth.
keypoints_3d - 4
Review
A person checks each episode. Rejected episodes never reach you.
quality report - 5
Deliver
LeRobot-ready dataset with a dataset card and the measured error.
dataset card
What's in every episode
The raw streams, the calibration that produced them, and the numbers you need to decide whether to trust it.
Spec sheet
ACT-RGBD-HAND3D-v1| Sensor | Orbbec Gemini 335, serial and firmware logged per episode |
| Streams | 1280×720 @ 30 fps clean RGB, raw uint16 depth (not resized), IMU |
| Timing | Device timestamps per stream, dropped frames reported |
| Calibration | Factory intrinsics and extrinsics, hashed and attached |
| Hand pose | 21 keypoints per hand, lifted to metric 3D with sensor depth |
| Accuracy | Estimated error in mm, measured against known references |
| Quality | Hand visibility, jitter, bone-length consistency, valid-depth coverage |
| Annotations | Subtask boundaries, success or failure, English instruction, objects and environment |
| Provenance | sensor_verified only with serial, calibration and checksums; demos are never mixed in |
Delivery
LeRobot-ready · in validationactem_workshop_pilot/ ├─ README.md # dataset card ├─ meta/info.json ├─ data/chunk-000/ ├─ videos/observation.images.rgb/ ├─ observation.depth/ └─ quality_report.md # error in mm
These are human demonstrations. Hand keypoints are not robot joint commands, and this is not teleoperation data.
Need a wrist view, a different task length or another format? Tell us in the form.
Where we capture
Most open hand datasets are tabletops and home kitchens. We record in workplaces where that data doesn't exist yet.
- WorkshopsFastening, sanding, small assembly
- PackingPicking, boxing, taping, labeling
- Commercial kitchensPrep, portioning, plating
- Flowers and coffeeCutting, bunching, sorting
- Garments and cablesFolding, routing, tying
What this data is, and who it's for
What this data is, and what it isn't
Human demonstrations, not robot actions
These are human demonstrations. Hand keypoints are not robot joint commands, and this is not teleoperation data. It's most useful for human-to-robot transfer, mid-training and evaluation alongside robot data you already have.
How we handle consent
Everyone recorded signs an explicit consent for AI training and international licensing. We capture hands and workspaces, not faces; faces, screens and plates are blurred in delivered video, and voice audio is removed. Consent is logged per episode and can be revoked.
Who we work with
Robotics labs and startups
Custom capture for the tasks your policy needs to learn, without building your own collection operation.
Data aggregators
Regional capture capacity in UTC-5 with our own 3D quality pipeline. We're open to working as a collection partner.
Tell us what you're training.
We're running early-access pilots with a small number of teams. Describe the task and we'll come back with a capture protocol and a few sample episodes to review.
A pilot starts by agreeing on tasks, objects, environments and acceptance criteria. You review the written protocol and sample episodes before full capture, every episode is checked by a person, and you get the dataset, its card and a quality report with the measured error. Pilots are scoped per project; we send a proposal after a short technical call.
Questions
What does a pilot cost?
It depends on tasks, volume and exclusivity. We send a scoped proposal after a technical call.
What format do you deliver?
LeRobot-ready delivery, currently in validation, with a dataset card. Other formats on request.
Can the data be exclusive to us?
Yes, custom captures can be exclusive. Tell us in the form.
Is this teleoperation data?
No. It's human demonstrations with 3D hand pose. See what this data is.
How do you handle consent and privacy?
Explicit consent per participant, no faces in delivered data, and processing under Colombian data protection law (Ley 1581 de 2012).
Where are you based?
Bogotá, Colombia. UTC-5, same working hours as the US East Coast.