Skip to content
Solutions · Physical AI & Robot Data

World Model Training Data: Real Video for World Foundation Models

World model training data captured from the real world: grounded egocentric video with synced metric depth, camera pose, action labels and full provenance. The physically faithful footage a world foundation model needs, not scraped mono clips. Delivered LeRobot-ready.

Book a World-Model Data Strategy Session Free 30-minute call · mutual NDA included
30HzSynced stereo RGB and metric depth per session.
200HzIMU on the head and both wrists — motion measured, not inferred.
8Ground-truth layers on every episode, synced on one clock.
Outcomes

What's in Our World-Model Dataset

Every episode is real, physically faithful capture: synchronized stereo video and metric depth, camera pose and ego-motion, action labels and provenance, far beyond the mono clips most video corpora ship.

Long Unbroken Takes

Continuous first-person footage, not chopped highlight clips, so dynamics stay intact.

Metric Depth, Not Just RGB

True-scale geometry per frame — the signal a world model needs to predict physics.

Real Physics Interactions

Contact, deformables and slip on real floors — the hardest things to simulate.

Camera Pose and Ego-Motion

Head and wrist IMU at 200Hz, so motion is measured, not inferred from pixels.

Action and Intent Labels

Verb-noun action segments and 21-point hand pose for action-conditioned training.

Clear Licence and Provenance

Signed consent and SHA256 chain of custody on every episode.

The Problem

Why Metric Depth, Not Just RGB

Monocular video is scale-ambiguous, so a model cannot recover true geometry from it. Our stereo capture ships metric depth per frame — the exact signal a world model needs to predict how a scene will move. Stereo remains underexplored in world and VLA models, which is where the gap is. Three things scraped or synthetic video structurally can't give you:

1 Scraped clips are mono and scale-ambiguous — no metric depth, no true geometry a world model can predict from.
2 Synthetic stalls on contact, deformables and slip — exactly where sim-to-real breaks for a physics-grounded model.
3 No consent or provenance — a licensing and audit liability the moment you post-train a commercial model.
The Real-Capture Answer

We ground the model on real physics, not inferred pixels.

Real video, metric depth and action labels on one clock — the grounded signal a world foundation model needs, and the one scraped mono video can't provide. Simulation is unbeatable for resets and perfect labels, but real capture grounds the model on the physics it will actually face. We treat sim and real as complements, not rivals.

Metric depth per frame, not scale-ambiguous mono
Real contact, deformables and slip on real floors
Consent and SHA256 provenance on every episode
Where It Ships

How You Receive and Use It

Every episode ships LeRobot-ready, with RLDS/TFDS and ROS 2 / Foxglove MCAP on request, so it drops into a Cosmos-style or Isaac Lab post-training workflow. One honest caveat: this is human egocentric capture — it grounds and post-trains a world model, and it does not replace your simulator or ship robot action labels out of the box.

1

Action-conditioned video prediction

Genie-style next-frame prediction, trained on real footage with real action labels — verb-noun segments and 21-point hand pose synced to the video.

2

Grounding Cosmos-style world models

Post-train an open world foundation model on the environments you actually care about, drawing on the same multimodal stack as our physical AI training data.

3

Scene geometry and 3D reconstruction

Metric depth plus ego-motion reconstructs scenes at true scale for occupancy and mapping — geometry a monocular corpus cannot recover.

4

Custom capture for your world model

Environments, objects and failure modes captured to spec, so the data you post-train on is your data, not a generic corpus everyone else already has.

5

LeRobot-ready, drops in without conversion

Every episode ships as a native LeRobotDataset, so it drops into a Cosmos-style or Isaac Lab post-training workflow without a conversion project.

6

RLDS/TFDS and ROS 2 / Foxglove MCAP

RLDS/TFDS and ROS 2 / Foxglove MCAP are available on request, so the data lands in the exact format your training and playback tooling already expects.

7

Grounds real physics where simulation breaks

Human egocentric capture grounds and post-trains your world model on friction, deformables and slip — exactly where sim-to-real breaks — without replacing your simulator or shipping robot action labels out of the box.

8

Representative sample pack to evaluate first

A small but fully representative delivery with all modalities and calibration, so your team can inspect the footage and sensor alignment before any commitment.

Start Today

Grounding a World Model on Real Footage?

Tell us the environments and dynamics you need. We'll map the capture and send a representative sample pack — the same modalities, calibration and delivery format you'd get in production, so your team can inspect the footage and sensor alignment before any commitment.

Book a Strategy Session →
Ask us about
Grounded egocentric video with synced metric depth and camera pose
Action-conditioned video prediction and Genie-style next-frame training
Post-training Cosmos-style world foundation models on real footage
LeRobot, RLDS/TFDS and ROS 2 / Foxglove MCAP delivery formats
Custom capture of the environments, dynamics and failure modes you need
Commercial licensing, consent releases and SHA256 provenance
Own the Capture

When You Need Custom Capture to Your Environments, Not a Generic Corpus

Scraped and synthetic video cover common footage well. But teams training real world foundation models need capture that a scraped or synthetic dataset structurally can't deliver:

Your exact environments, objects and dynamics — captured to spec, not the median scraped clip.
Metric depth measured first-hand per frame — not inferred from scale-ambiguous mono video.
Real physics where simulation breaks — contact, deformables and slip on real floors.
Ground truth on one synced clock — stereo RGB, depth, camera pose, IMU, action segments and hand pose.
Consent and provenance you can audit — signed releases and SHA256 chain of custody per episode.
Exclusive licensing on custom episodes — the data becomes a moat, not a shared corpus.

Send the environments, dynamics and edge cases you need and we scope a capture program against them — sites, objects, failure modes, episode counts and acceptance criteria — so the footage you post-train on is your data, measured on real floors, not a generic corpus everyone else already has.

Questions

Frequently Asked Questions

Every episode is captured under a signed release that permits commercial model training, with a per-episode chain of custody you can audit. Licensing is per-dataset and non-exclusive by default; exclusive and custom terms are available.

Native LeRobotDataset, plus RLDS/TFDS and ROS 2 / Foxglove MCAP on request. It drops into Cosmos-style or Isaac Lab post-training without a conversion project.

Scraped clips are mono and scale-ambiguous; synthetic stalls on contact and deformables. Ours is real, grounded, physically faithful capture with metric depth and action labels.

Yes. Send the environments, dynamics and edge cases you need and we scope a capture program against them, with exclusive licensing available on custom episodes.

Start with the sample pack — a small but fully representative delivery with all modalities and calibration, so your team can inspect it before any commitment.

See the Data Before You Decide

A 30-minute strategy call. We'll walk through the environments and dynamics you need, your target world model and delivery format — then scope a representative sample pack so your team can inspect the footage and sensor alignment before any commitment.

Discuss Your Project