Dataset · sample viewer
Spatiotemporal Gameplay
Real licensed gameplay, captured at the engine level — press play, then switch tabs to watch camera trajectory, keyboard, mouse, and depth.
Pristine gameplay at up to 4K/60fps, with HUD, menus, and UI stripped at the engine level before recording — models train on the raw 3D world, not screen capture. Long, uncut sessions preserve spatial layout and temporal dynamics across long horizons, and every recording ships with synchronized engine-level signals rather than pixels alone.
AI-generated scene labels across 1,000+ activity types — each segment tagged with what is happening, where, and when — alongside per-frame camera pose, raw input, and physics/game state as synchronized JSON.
Real, licensed gameplay captured with consent; HUD and UI removed at the engine level; no screen capture. Licensable for research use.
Pristine gameplay up to 4K/60fps. HUD, menus, and UI stripped at the engine level before recording — the raw 3D world, not screen capture.
In-game audio as separate tracks — dialogue, environmental sound, and effects at full fidelity.
Raw keyboard and mouse from hardware: exact key sequences, mouse deltas, and timing. Not inferred from video.
3D camera position, rotation, field of view, and velocity, extracted directly from the game engine.
Engine-level physics — object positions, collisions, velocities, and environmental state — as synchronized JSON.
AI-generated labels across 1,000+ activity types. Each segment tagged with what is happening, where, and when.
Sample Viewer
10 of 10 clipsThe clip you're viewing — pick another from the library below.
Sample 01
PLACEHOLDER — third-person traversal through a stylized environment.
- Frames
- 8,627
- Avg FPS
- 41.5
- FOV
- 85°
- Camera pose
- 6-DoF
- Input events
- 29,164
Camera pose (position + rotation), trajectory, and FOV — a first-class geometric signal for action control and long-horizon 3D consistency. Intrinsics / extrinsics available in the raw feed.
Browse