Multimodal Action World Model

This package demonstrates a VLM/VLA/world-model-style contract without introducing architecture-specific package kinds.

Scope

One artifact may provide semantic grounding, state estimation, dynamics/observation prediction, and action policy roles. OWP standardizes grounding, compatible World View/State Compiler contracts, interfaces, validity, lineage, and evaluation while leaving checkpoint format implementation-defined.

• World: openworld-examples/mobile-manipulation-world@0.1.0, with its views/pick-place-task.yaml View and state/pick-place-compiler.yaml State Compiler.
• Inputs: front and wrist RGB, language, proprioception, and action history, plus the EffectiveWorldState. Front RGB and proprioception are required.
• Outputs: latent state, predicted observation, action, and trajectory.
• Time: a context window of 8 steps, a prediction window of 4 steps, and an action horizon of 4 steps.

Sources

• openworld-examples/mobile-manipulation-world@0.1.0, a dependency, including its reference environment.
• models/model-artifact.yaml: the place to bind an immutable external model revision (Hugging Face, OCI, S3, or a local content-addressed store). No revision is bound.

Use it for

• Write down which World, View, State Compiler, and modalities a VLM/VLA model needs before you bind its weights.
• Predict the next observations and actions for the pick-and-place task, once a model is bound.
• Evaluate pick-and-place with a verified object pose change, not controller acceptance (eval/manipulation.yaml, eval/verifier.yaml).
• Tell current results from results produced under a superseded evaluation profile (see Evaluation lineage).

Limitations

• Example contract; no weights bundled.
• The metrics in eval/evidence-reference-sim.yaml are illustrative only; no evaluation was run.

Versions

• 0.1.0: first public example.

Evaluation lineage

eval/evidence-reference-sim.yaml binds a result to exact versions: EvaluationProfile manipulation-basic@0.2.0, Verifier pick-place-effect-verifier@0.1.0, the pick-place-task View, the pick-place-compiler State Compiler, and the reference environment. manipulation-basic@0.2.0 supersedes 0.1.0, so results produced under the older profile must not be read as current performance.

Assets

Path	Kind
eval/evidence-reference-sim.yaml	CompatibilityEvidence
eval/manipulation.yaml	EvaluationProfile
eval/verifier.yaml	VerifierProfile
models/adapter.yaml	RepresentationAdapterProfile
models/model-artifact.yaml	ModelArtifact