Gameplay data

Gameplay data for interactive world models

Capture synchronized controls, video, engine state, events, and rights evidence instead of another folder of clips.

Synthetic schema sampleObservation → action → next state
01
00.000sObservergb.player_view
02
00.050sActmove.forward + camera.delta
03
00.100sTransitionpose.x +0.84 m
04
00.150sOutcomeprogress 0.31 → 0.36

The values are synthetic and illustrate the public specification. They are not measurements from a commercial dataset.

Direct answer

Gameplay becomes useful world-model training data when player observations, controls, engine state, events, and outcomes are synchronized at episode level and released under documented rights.

Games expose structured dynamics

Instrumented games offer repeatable environments, rich action spaces, precise event boundaries, and measurable outcomes. Engine access can expose physics, object transforms, inventory, collisions, and task state that pixels cannot reliably recover.

Livestream video, spectator footage, bot matches, human demonstrations, and publisher-authorized private capture are different sources with different labels and rights.

Capture the complete control surface

Record keys, buttons, analog axes, pointer deltas, remapping, menu state, and pause state. Preserve raw input beside normalized controls so transformation choices remain auditable.

Loading screens, replays, camera cuts, menus, and disconnected sessions need explicit states. Treating them as ordinary play corrupts temporal learning.

  • Player-view video with stable timestamps
  • Raw device input and normalized action
  • Patch, map, build, graphics, and control configuration
  • Engine state, events, task progress, and terminal result

Commercial use starts with authority

Owning the recorder does not automatically clear a game's audiovisual content. Commercial supply requires documented authority for capture, model training, delivery, and customer use.

Public research releases can still be valuable references. Their inclusion in a directory must not imply that ActionTrajectories owns or resells them.

Questions this page should settle

What gameplay signals should be recorded?

At minimum, record observations, controls, episode boundaries, environment version, and outcomes. Engine state and events make the release substantially more useful.

Can public gameplay videos be sold as training data?

Public availability alone does not grant commercial training or redistribution rights. Each source and intended use needs a rights review.

Does ActionTrajectories offer publisher-authorized collection?

Custom programs can be scoped around the required title, instrumentation, collection method, rights, and delivery contract.

Bring us the training target.

We will map it to the observations, actions, state, outcomes, rights, and delivery evidence the dataset must contain.

Request access