For AI labs

The full state behind every frame.

Simulation-grade gameplay data for world models, agents, and robotics. Rights-cleared at the source, captured frame-perfect from the running game, and normalised across a catalogue of 50+ licensed titles. Ground truth, not estimated after the fact.

Captured directly from the running game with PRISM. Cleared to train, disclosable under the EU AI Act.
Ground truth

Read from the engine. Not estimated from pixels.

Most game data on the market is inferred: state read off the HUD with OCR, depth and normals generated after the fact by an estimator. That inference is also what generates the labels, under rules and definitions you never see. Every error poisons the dataset.

Ours is read directly from the running game. Labels are computed from that ground truth with explicit rules, through a shared game lexicon that normalises terminology across engines, transparently exposed and queryable.

Below, a real capture session with its signals playing in sync: scrub to any frame and read the exact state, input, and buffers behind it.

Real capture session: switch render buffers on the video, scrub to any frame and read the exact state behind it, all from the session's VTX file. Our homework, shown, not described.

Render

Rendering

Visual ground truth straight from the engine. Pixels and the render buffers behind them.

VideoDepthSegmentationSurface normalsMotion vectorsUI mask
Input

Player Inputs

What the human did, synchronised with the state stream. The action half of any imitation-learning pair.

MouseKeyboardControllerAction labelsSemantic keybindingsPlayer viewport
State

World State

Everything the game is doing, frame by frame. The structured world behind the screen, not reconstructed from pixels.

EntitiesTransformsPhysicsCollisionsHealth & statusEventsObjectivesRewardsCamera
Enrich

Data Enrichment

Added labels generated after capture. Semantic, inferred, and consistent across every supported title.

Scene captionsEvent annotationsIntent & trajectoryNamed actionsInferred inputs
+ added after capture
Clean RGB · HUD removed
UI mask · HUD isolated
Normals · surface direction
Depth · distance to camera
Normalised

One coordinate system. Every title.

Every title is normalised to a single coordinate system, scale, and convention. Engine quirks removed. A grant spanning dozens of titles behaves like one dataset, not dozens of integrations, so you train and mix across the whole catalogue with zero per-title wrangling.

RAGE Source 2 Unity Unreal Engine Hextech
Units 1 cm = 1 unit same scale, every title
Coordinates right-handed X right · Y up · Z forward
Rotations quaternions all engine formats converted
Transforms world-space local transforms resolved

Skeletal bone data: hierarchical, per-joint rotations, where cross-engine normalisation usually breaks. Each joint's rotation is relative to its parent, so a single convention error compounds down the entire chain, and no two engines rig the same way. We resolve all of it to one standard. Captured at run-time from real player sessions, not replays.

Licensed & disclosable

Cleared to train on. Ready to disclose.

From 2 August 2026 the EU AI Act obliges general-purpose AI providers to publish a summary of their training data; California's AB 2013 already requires disclosure. Gameplay cleared only at the player or platform layer is exactly what those summaries expose. Ours is licensed at the source, so you can name it and move on.

Studio + player cleared

Captured only from titles we license directly, with the right to train. The studio copyright layer, the one in every frame, is cleared, not just the player's recording.

Licence ID per session

Every session's manifest carries its source, licence, and rights scope. Full chain of custody, per delivery, not a blanket warranty clause.

Carve-outs tracked

Uncleared audio and third-party assets identified and handled per title, up front, not discovered later at diligence.

Synthetic capture doesn't escape this: data generated from a game is still derived from that game's IP. The rights question follows the title, not the capture method. Ours arrives already answered.

Work with it

Built for your pipeline.

A session is a folder you can open: standard mp4 for every visual signal, world state, inputs and events in VTX (our open, Apache-2.0 format), and one manifest describing it all. Install the Zenos Data CLI to search your grant, pull, and convert to Parquet or JSON. VTX spec on GitHub · CLI docs in the portal

the whole workflow · one tool
$ zenos grantsorder_0112 · 50,000 hrs · 38 titles · granted
$ zenos search --label driving --weather rain312 sessions · 1,840 hrs in your grant
$ zenos pull order_0112 --matchvideo + buffers + state + manifest
$ zenos convert ./order_0112 --parquetoptional · state to parquet
$ zenos view session_00428 --rerunframe-synced playback in Rerun
Re-label on demand

Your schema changes, the data doesn't. Re-derive labels from exact state, no recapture, no re-licensing. The data outlives any single training run.

Plain open formats

mp4 + VTX + one manifest per session. No proprietary reader between you and the data.

One licensed source

Every derivative traces back to the same cleared licence. The rights scope travels with the data.

How access works

Browse everything. Unlock what you license.

Request access and the catalogue opens in the lab portal: titles, labels, and hours up front, with free samples to evaluate straight away. The data itself unlocks per order. Agree a volume of hours across the titles you want, each IP owner signs off on the use, and the data lands in your workspace. From there it's self-serve: search at full depth, pull, train. Every session carries its licence ID and rights scope in the manifest.

01

Request access

Register in the portal. Free samples to evaluate immediately.

02

Scope

Agree hours, titles, and signals with us.

03

Signed off

Already licensed. Each IP owner confirms the use, built into the deal.

04

Granted

The data lands in your workspace.

05

Train

Search, pull, convert. Self-serve from here.

Build on real worlds.

Building frontier AI?Talk to us →
Have a catalogue?List your game →