PhysDBPhysical AI Map

Models

World foundation models

Models built to generate, reason about, or predict physical scenes and trajectories for training, testing, or planning.

world-modelgenerationreasoning

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

contains

Physical AI

Physical AI model taxonomy

One model family does not define the field.

generates context for

Synthetic data

World model training and evaluation

Synthetic scenes need domain-gap audits.

bridges to

Simulation

Model-assisted simulation

Generated world behavior is not physical proof.

supports

Embodied reasoning

Physical scene interpretation

Reasoning output still needs execution validation.

evaluates

Benchmarks

World model evaluation

Benchmark scope must be named.