Frame-by-Frame Prediction of Machine Dynamics
The model treats video as a physics readout. It ingests the current frames of a scene, conditions on a control input, and generates the frames that follow. Trained on footage of real machines, it learns dynamics that are hard to write down analytically: contact, load swing, the coupled motion of a heavy platform and the ground it stands on.
Because prediction is generative, the same model answers counterfactuals. Change the action and regenerate; the difference between rollouts is the cost of a decision, computed before the decision is made. That conditioning on control is what separates a world model from a video generator.
Planning, Simulation, and Training Data
A predictive model of the world earns its keep three ways:
- Planning. A controller proposes actions, the world model rolls each one forward, and the best predicted outcome wins. A vision-language-action policy can be scored the same way before it touches hardware.
- Simulation. Generated rollouts complement physics engines. A hand-built simulator encodes what engineers know; a world model encodes what the cameras saw. See simulation and real-to-sim.
- Training data. Predicted trajectories extend recorded datasets, giving reinforcement learning agents more scenes than any fleet could film.
Data acquisition closes the loop. Craneformerz, GENOMA's AI videogame and data-acquisition platform for robotic lifting, produces exactly the kind of machine-interaction data these models train on.
Deterministic Gates and Sovereign Serving
World models fail in ways a casual glance does not catch, so GENOMA checks them the way it checks all its models: verify by running. Every model generation passes deterministic gates before it ships, and a generation that fails a gate does not ship.
Training and inference run on GENOMA-owned hardware, following the local-first approach described on the sovereign model serving page. For everyone else, the robotic world models are coming to fal.ai. The launch is announced; the models are not live yet.
Questions
What is a robotic world model?
A generative model that predicts how a physical scene and the machine inside it evolve, frame by frame, conditioned on control inputs. It works as a learned simulator: propose an action, roll it forward, read the predicted outcome. GENOMA trains these models on footage of real heavy machines.
When will GENOMA's robotic world models be available?
They are coming to fal.ai. The launch is announced and the models are not yet live. GENOMA trains and serves them on hardware it owns, and each generation passes deterministic gates before release.
How does a world model differ from a VLA model?
A vision-language-action model maps observations and instructions to actions. A world model predicts the consequences of actions. They pair naturally: the world model rolls a candidate policy forward and shows what would happen before anything runs on a physical machine.