Live World Model Lab

An interactive UTMIST workshop demo that runs a real compact world model entirely in the browser.

The experiment trains five neural transition predictors from random initialization, uses their action-conditioned rollouts for model-predictive control, introduces a previously unseen dynamics change, measures the resulting prediction error, applies a fast residual update, and replans around the changed transition.

What is—and is not—being claimed

• Real: learned weights, held-out loss, ensemble predictions, uncertainty, rollout search, environment transitions, online adaptation, and replanning.
• Simplified: observations are compact occupancy grids and the predictors are small MLPs. This is not a pretrained video world foundation model.

That compact scope makes the complete causal loop visible, reproducible, and reliable during a workshop. The closing section identifies how to scale the same protocol to JEPA-WM, DINO-WM, or action-conditioned video diffusion.

Run

Requires Node.js 22.13 or later.

npm install
npm run dev

Open http://localhost:3000 and follow the single experiment button.

Validate

npm test

The numerical tests verify that the ensemble learns nominal dynamics, that MPC reaches the goal through the initially valid shortcut, and that a one-shot dynamics update causes replanning around the changed edge.