Video and actions

Action-conditioned video datasets make futures controllable

An action conditioned video dataset aligns frames with controls, state, timing, and outcomes so models can learn which decisions produce which futures.

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

An action-conditioned video dataset is a time-aligned collection of video observations, actions, and resulting state changes. The useful unit is an episode where each control input can be connected to the visual future it helped produce.

Video alone hides the control signal

The same visual context can lead to different futures when an agent turns, waits, accelerates, selects a tool, or issues a higher-level command. Passive frames show change, but they do not identify which control caused it.

A usable release keeps recorded actions separate from inferred actions. Estimated controls can add research value, but they are not measured ground truth and should retain model, confidence, and review metadata.

  • Frames, audio, depth, and observation timestamps
  • Raw and normalized controls with action semantics
  • Action issue time, effective time, and latency
  • State, events, progress, failure, and terminal outcome

Synchronization is the label

A control recorded after its visual consequence teaches the wrong causal order. Every stream should resolve to one declared clock with frame rate, action rate, capture delay, dropped-frame handling, and resampling method documented.

Commanded and executed actions can differ when a game, simulator, robot, or vehicle clips, blocks, delays, or overrides input. Preserve both when the environment exposes them.

Source type and rights survive packaging

First-party instrumented capture, publisher-authorized footage, public video with inferred actions, and generated or simulated episodes are different evidence classes. Combining them without source labels makes quality and rights impossible to audit.

Commercial supply requires authority for collection, model training, delivery, and customer use. Public research references can inform a specification without implying ownership or resale permission.

Questions this page should settle

What is an action conditioned video dataset?

It is a collection of video episodes aligned with the actions and state changes that produced each observed future.

Is passive video action conditioned?

No. Passive video shows visual change but does not directly record the control input responsible for a transition.

Do inferred actions count?

They can support research when clearly labeled, but they should retain inference provenance and uncertainty rather than be presented as recorded ground truth.

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