Emerging category

World action models make control explicit

Map an emerging model category to the observations, action labels, state, and evaluations its datasets must support.

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

A world model represents or predicts environment dynamics. A world action model makes actions a first-class conditioning or output variable so prediction can support control, policy learning, or interactive generation.

The name does not define the interface

World-model architectures differ in what they observe, predict, and expose to a policy. Some condition on measured actions. Others learn from passive sequences and provide no direct control surface.

World action model is an emerging label. The useful technical question is whether actions condition the future, are predicted jointly, or serve as policy outputs.

Action trajectories are the data layer

A trajectory connects observation history, control input, state transition, and outcome without assuming one model architecture. It can support controllable video, behavior cloning, offline reinforcement learning, and learned simulation.

Each application still needs a specific action ontology, temporal resolution, environment coverage, and evaluation split.

Avoid category theater

Renaming passive clips as action data does not make a model controllable. Measured, inferred, generated, and unavailable actions must remain distinct.

Evaluate intervention sensitivity: change the action while holding context fixed, then measure whether the predicted future changes in a plausible and timely way.

Questions this page should settle

Is a world action model the same as a world model?

The terms overlap, but world action model usually emphasizes a system that represents, predicts, or conditions on action explicitly.

What data trains a world action model?

The core requirement is temporally aligned observations and actions. State, outcomes, and environment metadata improve training and evaluation.

Why call the company ActionTrajectories?

The name describes the model-agnostic data primitive: sequences connecting what an agent saw, did, and caused.

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