Genie 4 puts Gemini in charge of the simulation as its planner, and the humanoid playing checkers is the demo I'd watch.

Google calls it its most realistic and interactive world model yet. You steer the scene in real time with plain language: weather, camera angle, new objects, actions.

The interesting part is the loop. Per Google, Gemini acts as the brain. It turns your request into actions, plans multi-step events, watches the scene unfold and adjusts when you change the story. One summary names Gemini 3.8 Flash as the model that watches the scene, breaks a task into steps and checks whether each step worked.

From official announcement:
> native audio matched to the world, including characters who can talk in real time
> photorealistic scenes with real-world physics

On the robotics side, DeepMind reportedly showed an experiment pairing Gemini 4 Argon with Genie 4 to simulate an Apollo humanoid playing checkers. One commenter's idea for a follow-up: move an object mid-task, let Gemini replan, see if the agent recovers. That is a test someone should run, and I'd want to see the numbers.

Access today is a research preview for trusted testers. Google AI Ultra subscribers get it soon through Project Genie, and early access partners via the Gemini API and AI Studio. No dates given.

bookmark this and check the Gemini API docs when early access opens, so you can try the replanning loop on your own task