The episode is the useful unit
Loose clips can teach visual priors. An episode preserves the initial condition, action history, environment state, task context, and terminal result needed to learn dynamics.
Each stream should resolve to one declared clock. Sampling rate, dropped frames, action delay, and uncertainty belong in the release record, not in private implementation notes.
- Observed pixels, audio, depth, language, or sensors
- Raw and normalized action signals
- Measured engine or simulator state
- Events, progress, reward, failure, and terminal state
Hours do not describe coverage
Nominal duration can hide repetition. A useful release reports coverage across environments, task states, action combinations, outcomes, rare events, failures, and recoveries.
Define the coverage matrix before collection. It becomes the contract connecting the research target to collection quotas and release acceptance.
Inspect the public structure before discussing scale
Our JSON Schema declares required episode, timebase, record, provenance, rights, and quality fields. The synthetic example shows nanosecond timestamps, raw and normalized controls, action effective time, state, progress, and validity masks in one episode.
These files are proof of the documented structure, not proof of available commercial recordings or training results. All example values are synthetic; media paths are placeholders, hashes can be null, and no accuracy or coverage claim should be inferred from the example.
Require release evidence before accepting a dataset
For a scoped commercial delivery, agree the evidence package before collection or licensing. A representative authorized sample should load through the intended training pipeline. Timing, missing data, source identity, rights, and split boundaries should remain inspectable after export.
Availability, volume, source access, formats, price, and licensing are confirmed per proposal. Public reference listings are not our commercial inventory. A synthetic example cannot replace a release-specific sample or quality report.
- Versioned manifest with episode IDs, stream counts, file paths, and checksums
- Timing report with clock, sampling rates, delay, drift, and missing-data handling
- Coverage summary by task, environment, action, outcome, and failure class
- Source and rights record identifying permitted uses and restrictions
- Loader acceptance results and train/evaluation split lineage
Choose data for the model's actual objective
Spatial reconstruction, video prediction, and action-conditioned dynamics need different evidence. The announced AMD and World Labs deal is a reason to clarify the learning target, not a reason to treat all spatial-AI data as executable action trajectories.
Start with what the model must predict or control. Then specify observation history, action interface, state fields, evaluation horizon, source classes, and acceptance criteria.
Rights are part of the dataset
Download access is not commercial permission. The release needs a traceable source, rights holder, allowed uses, redistribution terms, attribution requirements, restrictions, and withdrawal path.
ActionTrajectories keeps external research references separate from datasets commissioned or supplied under a commercial agreement.