Contact labels. With their error.
Labels you can check.
In formats you read.
Video shows the motion. The kit is to measure the contact, and the card is to say how well.
Contact labels with their error.
Per-finger contact force from bare hands, calibrated against a fingertip-force rig. The card is to print the error on people held out of training and the measured sync error. A contact label is released only when its card prints both numbers as measured values.
An audit report.
The open contact-truth audit is to score the labels against rig ground truth: held-out-user error, sync error and coverage. It is to ship with the harness and the rig reference episodes, so your team can reproduce each number.
Delivery that your stack reads.
LeRobot v2.1, LeRobot v3.0 is written and validated in the synthetic demo. is planned. Each export carries its dataset card. The MCAP source of record of each episode stays on the capture hub and is never delivered.
Consent lineage for each episode.
Each episode is to carry its consent scope and its revocation state. The exporter refuses an episode after a withdrawal, and the licence passes the revocation duty to you.
Trust a measurement.
Not a promise.
Before a pilot, a team needs evidence that it can check. Each part below is a plan. The gates and the kill criterion stay in the registry, and the page does not copy their numbers.
A pre-registered test.
A force-label test and a force-aware against motion-only policy ablation, with the pass rule and the kill criterion written before the first recording. The result is to be published if it passes or fails.
Held-out evaluation on your robot.
You choose a task and hold it back. You run both policies on your own robot, on one sealed and interleaved schedule. The acceptance metric is a policy result on your task, not hours delivered.
Delivery into your storage.
A signed export into storage that you own, or capture at your site. Raw sEMG never leaves the device, and video and poses leave only after redaction and under current consent.
A security posture plan.
Cr4dle holds no security certification today. The plan puts SOC 2 Type I first, then Type II. The start date is not set.
Licence terms.
The data is licensed, not assigned, for training and evaluation, with no resale. A licence never gives purposes broader than the consent of each person.
Where the hand goes next.
When contact starts.
A humanoid that works near people needs a model of human motion within its reach. The research is a human-behaviour layer: hand pose, head pose, gaze target and contact force from people, with kinematic next-action and reach-target labels that can condition an open world model. A separate research item is a sealed, held-out evaluation of robots near people, which needs its own protocol and decision.
This is research, not an offer. The evaluation is not a certification and not conformity evidence. The ROS 2 package gives inputs to speed-and-separation monitoring, and it never commands an actuator. Each simulation number is labelled as simulation.
What exists today.
What does not.
The capture path runs end to end on synthetic data and the simulator. The adapters, the calibration and the redaction stage are tested reference code. No session on real sensors is recorded, and no sensing performance is measured.
What is planned for a later phase: collections of consented, labelled episodes on your task (quoted per usable demonstration), the kit licence so your own demonstrators produce them, and dataset licences — each behind its dated gate. Cr4dle is a non-commercial research programme now, and nothing is offered for sale today.
Bring a task.
Define the proof.
Tell us the task, the robot, the policy that you compare against and the result that would make the data useful. Robot access is not needed for this conversation.