We license games from their rights holders, have real people play to your brief, and read depth, camera pose, events and state from the running game while they play. Every layer ships time-aligned to your AWS S3 or GCP bucket.
Depth and pose are estimated from pixels, which leads to visual drift and spatial hallucinations.
Depth comes from the engine’s depth buffer and camera pose from game memory. If a memory read fails, we fall back to visual SLAM and label that pose as estimated.
No paired action data or input logs for VLA model conditioning.
Keys, mouse and controller as pressed, time-aligned with every frame and every co-op view.
Edited clips and highlight reels, with no task and no record of what the player was trying to do.
Players get a task before they act. We keep failed attempts and recovery, not just clean play.
Scraped internet media carries severe legal and auditing risks.
Each game is licensed by contract with its rights holder before we record it.
Both come from the same licensed games and studio partners. Pick the one that fits how your team learns from worlds, or combine them.
Real players play licensed games to your brief. You get every layer, time-aligned, delivered to your bucket.
A licensed, realistic game hosted privately for your lab, ready for your agents to play from day one.
MP4 + Parquet, one file set per session. Sync is checked on each game. Each layer ships as its own file, because a recorded click is not an outcome label.
| Layer | What it is | Source |
|---|---|---|
| RGB, as displayed | Game window capture | Captured |
| Co-op views | Up to 4 first-person views, one clock | Captured |
| Player inputs | Keys, mouse, controller as pressed | Captured |
| Game audio | From the game process | Captured |
| Clean frame | Captured before the interface is drawn | Render buffer |
| Depth | Depth buffer, validated per game | Render buffer |
| Camera pose | Camera transform read from memory | Game memory |
| Events + state | Health, items, positions, read live | Game memory |
| Pose, estimated | Visual SLAM when the memory read fails | Estimated |
| Tasks + outcomes | Instruction, attempt, success, failure | Reviewed |
Render buffer and Game memory values are read from the running game, with no source code needed. We confirm coverage per game before you license.
We come from the games industry. We find the games, run the sessions and label the data around your research. Sessions ship as MP4 + Parquet by default, and we restructure them to your data loader’s exact spec when you need it.
We are onboarding 5 to 10 frontier AI research labs for our Q4 Design Partner Cohort. As a design partner, your team sets the games, tasks, output schemas and frame rates. We source and license the games, brief the players, and deliver evaluation datasets to you.
5 to 10 labs per cohort.
Small by design so each lab gets dedicated time from our team.
Four quick fields. We reply with a sample schema and a time for a demo.
Thanks. We will send a sample schema and propose a demo time.
