REC episode 04 182 t 00.000 s 0 / 5 channels 500 Hz · one clock

Training data for humanoids

We're building the richest training data for humanoids.

We capture the full kinetic chain of human movement, with the muscle activation and contact forces a camera can't see, and map it to humanoid joints.

Request a data briefing Scroll to start the capture ↓

Channel 01 · Full-body motion

Every joint, fingertip to floor.

A camera gets you this far: where the body is, 500 times a second. It's the part everyone already has.

Channel 02 · Muscle activation

The effort behind the motion.

Muscles fire before anything visibly moves. We record that effort along the whole chain, so a robot learns the intent, not just the result.

Channel 03 · Fingertip pressure and slip

How hard the hand holds on.

Pressure under each fingertip, shear across the pad, and the moment an object starts to slide. This is where grips crush and cups drop.

Channel 04 · Ground reaction

The floor pushes back.

A reach starts in the feet. We measure how they load and push, so the robot's balance is learned from a body that never falls over.

Channel 05 · Balance

Weight, kept over the feet.

The centre of mass shifts to carry the load. Together, the five channels make one episode: synchronised, labelled and ready to train on.

Output · Joint-mapped episode

Mapped to the humanoid.

Each channel is retargeted to the matching joint of the robot you train. Import it, run it, compare it against your own policy.

See the episode format

The gap

Same lift. Two very different records.

Most humanoid training data is video and teleoperation: where the body went, nothing about the forces that got it there. The robot has to guess the rest. That's where it fails.

Camera / teleop

Joint positions

Muscle activationnot captured

Fingertip pressurenot captured

Slipnot captured

Ground reactionnot captured

Balanceinferred

Hyletic episode

Joint positions

Muscle activation

Fingertip pressure

Slip

Ground reaction

Balance

Channels

Seven streams. One clock.

Everything is sampled together and time-aligned at source, so a fingertip event and a foot event in the same episode share the same timestamp.

shared clockt 00.000 scontact event at t 2.41 s

Channel	What it measures	At cursor	Rate
01	Full-body motionJoint positions and velocities	–°	500 Hz
02	Muscle activationEffort in the fibres behind each move	–%	1 kHz
03	Fingertip pressureNormal force at each contact	–N	500 Hz
04	Shear and slipTangential force and slip onset	–N	500 Hz
05	Ground reactionForce under each foot	–N	500 Hz
06	BalanceCentre of mass over the base of support	–m	500 Hz
07	Object stateMass, shape, pose	–	per episode

Drag across the traces to scrub.

Episode

What a delivered episode looks like.

One interaction, from first intent to release. Synchronised, calibrated, labelled, with usage rights attached, and retargeted to the joints of the humanoid you name.

Synchronised at source, not aligned afterwards.
Labelled with task, object, outcome and failure events.
Joint-mapped to your robot's kinematic model.
Rights-clear: every participant consents to training use.
Formats: LeRobot and RLDS, or your own schema.

// episode manifest (illustrative)
{
  "episode":  "hy-04182",
  "task":     "reach_and_lift",
  "object":   { "kind": "cup", "mass_kg": 0.31 },
  "duration_s": 6.84,
  "clock":    "monotonic, shared",
  "channels": [
    { "id": "motion",   "rate_hz": 500,  "joints": 34 },
    { "id": "emg",      "rate_hz": 1000, "sites":  16 },
    { "id": "fingertip","rate_hz": 500,  "pads":   10 },
    { "id": "grf",      "rate_hz": 500,  "plates": 2 },
    { "id": "com",      "rate_hz": 500,  "derived": true }
  ],
  "events":   [ { "t": 2.41, "type": "contact" }, { "t": 4.93, "type": "slip_margin_min" } ],
  "retarget": { "robot": "<your humanoid>", "dof": 28 },
  "rights":   "training, consented"
}

Who we record for

Who we record for.

Humanoid makers

Your robot walks. Picking things up without crushing or dropping them is the last mile. Full-body episodes with contact forces, retargeted to your joints.

Robot foundation-model labs

Video and teleop data scale, but they're blind to force. We add the channels the model has never seen: muscle activation, fingertip pressure, slip, ground reaction, all on one clock.

Dexterous-hand and tactile teams

The hardest signal is the moment before a slip. Per-fingertip pressure and shear with the labelled event, from real human hands.

Simulation and world-model teams

Contact is where simulators lie. Ground-truth forces to calibrate against, episode by episode.

Our prediction

By 2040, humanoids will move like us. Indistinguishable, apart from the mandatory regulatory tag.

The bodies are arriving. We capture more of the human than robots can use today, so the data is waiting when they do.

20262029203220362040

Projected gait figures for illustration. Not measured data.

2026 Hard shells and motors. Robots walk, stiffly.

2029 Lighter frames, better hands.

2032 Compliant bodies that feel contact.

2036 Synthetic muscle. Force placed at the right moment.

2040 Movement indistinguishable from ours. Tag H-2040.

Log

Always recording.

What we shipped, tested and attended. Newest first.

2026-10-02
NVIDIA GTC Berlin. We'll be at the humanoid sessions and the meetups around them. If you're training a robot, find us.
2026-10-01
Site live. Hyletic AI goes public with the capture programme.

Contact

Request a data briefing.

Tell us what you train and what keeps failing. We'll walk you through the channels, the episode format and how it maps to your robot. Investors: same form, we read everything. Academic lab? Ask about research access.