Build the worlds your model trains in.

Physicl is the platform physical AI teams generate their training data on. Describe a scene, vary it millions of ways, and get the ground truth with every frame.

Request access → Talk to us

Studio, our first app, is in private beta

Segmentation

Normals

Depth

Physicl Studio · warehouse DC-04 · drone run 017 · 342 frames at 24 fps

Working with teams building physical AI

More footage is not
what is missing.

Recorded · one office floor · 5 seconds at 24 fps

0 frames, and still nothing attached.

An hour of this is 86,400 frames, and not one of them carries depth, normals or an instance ID. Another hour of the same floor adds no new situations either.

Studio · private beta

Introducing
Physicl Studio.
The end of the fixed dataset.

Studio finds the scene your model has never seen, varies it on the axis you choose, and writes the ground truth onto every frame it renders.

Request access → How it works →

10k+scenes to start from

100k+objects to place

Our warehouse robot learned one aisle. It stalls as soon as the racking is restocked.

Searching 10k+ scenes · 100k+ objects

Then vary the aisle, not the building. Same racking, pallets in other places, light from midday through to night, labelled on the way out.

Which ground truth do you want on every frame?

Depth metres, EXR Surface normals RGB Segmentation + instance IDs Camera pose per frame 2D boxes + classes

480 runs of this aisle. Move the pallets, change the density, take the light through the day.

480 runs · 4.0 hours · 14.2 GB

cam 011080p · 24 fps

Rendered frame of a warehouse aisle, blue and orange pallet racking with boxes — Frame

Depth pass of the same aisle, near surfaces bright and far surfaces dark — Depth

Surface normals pass of the same aisle, each surface coloured by the direction it faces — Normals

Instance segmentation of the same aisle, every object filled with one flat colour — Instances

The same aisle with a 2D box and a class name drawn on every object — Boxes

Rendering 480 clips0 / 480Layout

One scene, every version of it

One scene,
millions of versions.
You choose what changes.

Each row moves one thing on the same scene: the camera, the layout or the light. Everything else stays where it was, so two runs can be compared directly.

Camera

orbit, 0 deg

robot height, 45 cm

egocentric, head mounted camera at 1.6 m

egocentric, 1.6 m

orbit, 120 deg

crane, start

crane, 2.2 m

lateral track

24 mm, wide

Layout

desks rotated

pods in the walkway

corridor obstructed

pallet, ladder, trolley

dense benching

planting between banks

floor stripped back

desks to the walls

Light

midday

late afternoon

golden, raking

golden

blue hour

dusk, lamps on

night, ceiling only

overcast

Output

beauty

depth, metres

depth, close range

surface normals

segmentation

instance IDs

camera pose

all layers, stacked

Who it is for

Built for the teams teaching machines the physical world.

Robot learning

I train robot policies.

I need the shelf restocked, the light changed and the clutter moved before my robot meets it for real.

Robot Learning Engineer
Research Scientist, Manipulation
Embodied AI Engineer

→ variations of one scene

JLPerception

I build vision models.

I need depth, normals and instance masks that are exact, without an annotation queue.

Perception Engineer
Computer Vision Researcher
3D Vision Engineer

→ ground truth on every frame

World models

I train world models.

I need varied, physically consistent video, with the camera known on every frame.

World Model Researcher
Research Scientist, Video Generation
Simulation Engineer

→ hours of footage per run

MKEvaluation

I test models before they ship.

I need scenes my model has never seen, and to rerun the same ones on every checkpoint.

ML Evaluation Engineer
Benchmark Lead
Robotics Test Engineer

→ a repeatable benchmark

From the render

Measured,
not estimated.

Nothing here was labelled by hand. The scene exists before the frame does, so depth, normals and material indices come out of the geometry with it.

01BeautyThe frame your model sees. Up to 4K, 24 fps.

02DepthMetres, EXR, per pixel. Read off the geometry.

03Surface normalsCamera space, one direction per pixel. EXR, or RGB-encoded PNG.

04Material indexOne index per pixel, so you can supervise surface by surface.

05Instances and boxesOne ID per object, stable across the run. 2D and 3D boxes, pose, visibility.

A kitchen rendered by Physicl Studio, marble island, wooden cabinets and three pendant lamps

Beauty

Depth

Surface normal map of the same kitchen, each surface coloured by the direction it faces

Normals

Material index of the same kitchen, every material filled with one flat colour

Materials

One run

What lands
in your bucket.

One run out of a batch of 480, as it arrives. Every file was written by the render.

run 042 of 48030 s · 24 fps · 720 frames

run_042/

frames/720 × png, 1920×1080, sRGB

shading/720 × png + exr, albedo, attenuation, trimap

depth/720 × exr, metres, 32 bit float

+ relative and colour-mapped png, fixed range across the run

normals/720 × exr, camera space, signed vectors

+ OpenGL, DirectX and OpenCV conventions, png or exr

+ world space on request

instances/720 × exr, one integer ID per annotated object, stable across the run

+ one colour per object, png

materials/720 × exr, one integer ID per material, stable across the run

+ one colour per material, png

uv/720 × png + exr, U and V per pixel, 32 bit float

annotations/720 × json, camera, 2D and 3D boxes, pose, visibility, per frame

+ boxes drawn on every frame, png

*.mp4one video per picture pass

Build the worlds your model trains in.

More footage is not
what is missing.

Introducing
Physicl Studio.
The end of the fixed dataset.

One scene,
millions of versions.
You choose what changes.