Taught by people.
Watch Fen clear a breakfast table, sort a basket of laundry, and pour a glass of water — one uncut take, no script, no operator.
Fen learns without a joystick
The real constraint in robotics is data. A single household task can take thousands of demonstrations, and the usual way of gathering them — teleoperation — is slow, expensive, and intrusive. Our bet runs the other way: autonomy that scales with people instead of operators.
That is what the Skillprint Cuff is for. It lets our Motion Partners produce training data at ordinary human speed, long before a robot is in the room. Fen’s hand was drawn as a mirror of the cuff — same joint spacing, same sensor placement — so a skill recorded on a wrist transfers to the machine without translation.
Data capture
Total episodes: 8.4MWe have shipped 3,180 cuffs to Motion Partners so far. They turn the parts of a day nobody films — reaching past a cabinet door, steadying a bowl, wiping, re-gripping a handle that slipped — into training data.
The dataset is the whole advantage. Cleaner demonstrations translate directly into faster skill acquisition and calmer motion. Fen is not following a script: it has watched enough people do the thing to have a view on how it is usually done, and what to do when it goes wrong.
From recording to behaviour
We recruit slowly and on purpose. Motion Partners are households, not contractors in a warehouse — chosen so the dataset covers galley kitchens, shared houses, and rooms with children, dogs, and clutter already in them.
One system carries every step: recruiting partners, capturing motion, training the model, scoring the result. A partner performs a briefed action while the cuff records force, joint angle, and two camera views at 120 Hz.
Each cycle inherits the last. Our evaluation team reads the failures first — the drops, the second grabs, the half-second hesitations — and rewrites the next week’s briefs around them. Recovery, not perfection, is what makes a robot usable in a real room.
Specifications
Fen is compliant by construction. Push it and it yields; cut power at any joint angle and it settles instead of falling. Nothing in the arm can hold a position it could not hold unpowered.
Collision avoidance runs against static and moving obstacles. Fen only attempts behaviours we have explicitly taught and reviewed — there is no exploratory mode running in your house.
Because training data comes from Motion Partners, we never need footage from your home to improve the model. Sharing a clip of a failure is opt-in, one clip at a time, and revocable.
A component that converts electrical, hydraulic, or pneumatic energy into motion, allowing a joint or mechanism to move.
Control of a robot by a human at a distance. The operator sees what the robot sees and sends commands directly to its arms, wheels, and hands.
Our wrist-worn recorder. It is a one-to-one representation of Fen’s hand, so anything you can do wearing it, Fen can be taught.
A control strategy that lets a robot yield, flex, or adapt its motion in response to outside force, instead of behaving like a rigid, unstoppable machine.
Our growing dataset of in-the-wild demonstrations, recorded by Motion Partners with Skillprint Cuffs. As of March 2026 it holds roughly 8.4 million episodes.
A system is passively stable if it settles into equilibrium through its own mechanical design, without ongoing actuation to hold a configuration.
A household that records demonstrations with our cuffs. Partners represent homes across eleven states and collect in the rooms they actually live in, not in a warehouse.
When a robot acts on its own — sensing the room and completing a task without moment-to-moment guidance. The opposite of teleoperation.
The robot foundation model we train in-house at Alder. It learns from data captured by Motion Partners wearing Skillprint Cuffs.