Building the data infrastructure for physical intelligence.

The human world is the training set. We capture high-fidelity human motion and interaction in the real world — and turn it into robot-ready data.

Modality

Full-body · Hands · Ego

Signal

6-DoF trajectories

Environment

Unconstrained, in the wild

Team

Three Stanford graduates

The problem

Language models had the internet. Robots have almost nothing.

Physical intelligence is bottlenecked by data, not architecture. Robot learning needs enormous quantities of high-quality real-world interaction — captured with human dexterity, at human scale, in unconstrained environments.

Text corpora

Trillions of tokens

Scraped, cheap, effectively unlimited.

Robot data

Orders of magnitude short

Teleoperation is slow, costly and lab-bound.

Human motion

Continuous, everywhere

Every hand, every day — currently uncaptured.

Our system

Wearable hardware that turns a person into a capture rig.

A coordinated set of worn sensors records motion and interaction while the wearer simply goes about real tasks.

Head-mounted

Egocentric vision and head pose. What the hands do, seen the way a robot will see it.

Tracking gloves

Finger articulation and hand pose at fidelity sufficient for dexterous manipulation.

Wearable cuffs

Arms, legs and torso. Lightweight sensing that disappears into ordinary clothing.

Mobile capture

No studio, no markers, no volume. Data collection wherever work already happens.

From humans to robots

Human movement → structured motion data → robot training.

01. Human

A person performs an ordinary real-world task.

02. Tracked skeleton

Joints, fingers and 6-DoF trajectories are resolved in real time.

03. Structured data

Motion becomes aligned, machine-readable interaction sequences.

04. Robot

Sequences train policies for humanoids and manipulation systems.