What it is
World models sit between perception, simulation, data generation, and policy learning. Cosmos 3 is one current example in the source registry.
Why it matters
They can make physical hypotheses and training data cheaper to explore, while still requiring validation against real measurements.
How not to overread it
Generated worlds are not evidence that a robot can safely perform the task in reality.
Related edges
Physical AI
Physical AI model taxonomy
One model family does not define the field.
Synthetic data
World model training and evaluation
Synthetic scenes need domain-gap audits.
Simulation
Model-assisted simulation
Generated world behavior is not physical proof.
Embodied reasoning
Physical scene interpretation
Reasoning output still needs execution validation.
Benchmarks
World model evaluation
Benchmark scope must be named.