SHOW ITHOW.
HUMAN DEMONSTRATIONS
FOR ROBOT LEARNING.
Recorded motion is playing. Try it loads the local 3D station and contact physics. Camera stays off until you choose it.
INFINITE ORDINARY.A KITCHEN. A HAND. ONE VERY LEARNABLE MOMENT.
YOUR HAND.
ITS NEXT
MOVE.
A small human thing. A big step for a robot. Human demonstrations for robot learning, captured in the places where life actually happens. Start with this mug.
Show your palm. Bring your thumb and index finger together. The robot hand closes around the mug; lift your hand and the mug follows the contact. Open your fingers and gravity gets it back.
This is the smallest useful demonstration: an intention, a trajectory, an object interaction, and an outcome. Replaykin turns many of these ordinary moments into data your robotics team can actually inspect.
Your browser will ask for camera permission. Vision and physics load only after your tap. Nothing is uploaded.
A MOTION ISN'T
DATA UNTIL
YOU CAN READ IT.
Scrub the take. Find the approach, the first contact, the lift, and the release. Each frame carries a timestamp, the hand command, object pose, and the contact-derived action label. Tracked frames carry 21 hand landmarks; missing observations stay empty. The sample and your own take use the same viewer and export format.
When a hand disappears behind an object, that gap belongs in the record. Camera frames carry a source, tracking state, and observation age. Missing detections have null keypoints; observations older than 800 milliseconds are marked stale and also exported as null. The included sample and control-generated points are explicitly labeled synthetic. Your model should know what the capture actually saw.
One inspectable take.
Included sample: 12 seconds, recorded at 30 Hz from the same contact simulation. Reach, grasp, transport, and release remain separate moments.
YOUR ROBOT.
YOUR STACK.
MORE REAL LIFE.
Replaykin is for teams training humanoids, mobile manipulators, and dexterous hands. Start with a task family your policy struggles with: handling cups, clearing a counter, sorting everyday objects, or recovering from a failed reach.
We design a collection around variation that matters: different hands, object shapes, lighting, viewpoints, and approaches. A hundred almost-identical takes from one person do not count as meaningful diversity.
The browser take demonstrates the shape of an episode. Production video, multi-view geometry, calibrated depth, and robot-specific retargeting are scoped separately. Tell us your embodiment and evaluation task before we collect. We would rather reject a poor match than sell you unusable hours.
CONSENT IS
PART OF
THE EPISODE.
People know what they are capturing, who can use it, and how they are paid. Production collection uses approved tasks, signed permissions, and a review before delivery. Private screens, bystanders, addresses, and personal documents do not belong in a training set.
In this browser, your camera is optional. There is no background upload, no audio capture, and no account. Stop the camera at any time. Raw video is never put in the JSON; your exported file contains numeric motion and interaction data.
FRAME PASSES.
Blur, lost tracking, an unclear outcome, or missing consent can reject a take. The acceptance rules are agreed before capture. You see rejection reasons, coverage gaps, and known limitations in the delivery report.
PAY FOR
ACCEPTED
MINUTES.
Not every minute recorded is a minute worth training on. Price your first collection by accepted data, with the task design and review rules agreed in advance.
One program setup. One transparent rate. Rejected takes do not quietly become billable hours.
USD, one collection. Includes the task spec, contributor coordination, standard hand-pose labels, quality review, and versioned delivery.
Start with 20 accepted hours for $5,760.Choose one task family and up to five environment variations. Review the first two accepted hours before the remaining collection proceeds. The pilot has no automatic renewal.
Outside that price:custom capture hardware, calibrated depth, specialist locations, exclusive licensing, and embodiment-specific retargeting. Your quote names these separately. If the acceptance criteria change, the scope changes before the invoice does.
GIVE IT ONE THING
WORTH LEARNING.
Bring a failure case, not a shopping list of hours. We will turn it into a collection spec that your learning team can evaluate.
Name the task, embodiment, success signal, required observations, and unacceptable gaps. Agree on data rights and the acceptance test.
Review the first two hours together. Check object coverage, hand visibility, annotation consistency, and whether the samples match the problem you actually have.
Keep a holdout set. Compare against your existing data. Expand the collection only when the first delivery earns its place in your training run.
Your technical lead can leave with something concrete: the sample episode, the schema, and a pilot brief with the current price and scope.