WorldAtlas
Summary
Section titled “Summary”WorldAtlas evaluates generative world models along a graded capability hierarchy—from perception and persistent scene modeling through manipulation, dynamics, and closed-loop interaction—rather than by surface fidelity alone.
Capability Gradient
Section titled “Capability Gradient”World models are often judged by the visual plausibility of their outputs, a criterion that can obscure sharp differences in underlying capability. A more informative view arranges these capabilities along a graded hierarchy:
- Perception — recovering information already present in an observation
- Modeling — maintaining a persistent internal state beyond any single frame
- Manipulation — transforming that state under an explicit constraint
- Dynamics — propagating state through time under physical and causal law
- Interaction — producing action-conditioned, counterfactually valid rollouts
Competence at one level does not entail competence at the levels above it.
WorldAtlas evaluates models against this gradient rather than against surface fidelity alone: its suites and metrics are designed to identify which level of capability a given system actually exhibits, from perception and scene modeling through physically consistent dynamics and closed-loop interaction. The repository mirrors this separation of concerns, with evaluation logic, adapter contracts, third-party checkouts, and cached checkpoints organized to evolve independently.
Documentation Map
Section titled “Documentation Map”Install, data layout, and worldatlas CLI: Getting Started. For the formal benchmark framing and task decomposition, start with the Benchmark Overview. The rest of the documentation set is organized as follows: