Google DeepMind unveils Genie 4 world model with audio, pitches it for robot simulation

Google DeepMind said in a blog post on Oct 27 that Genie 4, its new real-time world model, is rolling out as a research preview to trusted testers, with Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio next. Genie 4 builds on Gemini Omni for appearance and behaviour and is the first Genie model to generate audio natively, including characters that speak and react to user input.

For robotics the relevant part is control. Gemini 3.8 Flash runs in the loop as a guiding agent: it watches frames as Genie renders them, splits a goal into steps and checks whether each step succeeded. In one demonstration Gemini 4 Argon broke a task into steps that were fed to Genie 4 as text captions while it simulated Apptronik's Apollo humanoid playing checkers; another prompt had a Universal Robots UR5 arm dispensing liquid with a pipette. DeepMind calls such worlds a source of training environments for robots and agents and says enterprise partners are testing Genie 4 for scenario generation.

The limits are stated plainly in the post. Detail and object permanence drift as a session runs longer or after several prompt edits, and complex physical interactions, spatial accuracy and long-term consistency remain open research. The model page puts the useful window at a few minutes of continuous interaction.

MEKAWIRE context: Genie 4 lands one day before Google Cloud put Gemini Robotics ER 2 into enterprise preview and two weeks after Gemini 4 Argon for Robotics, so Google now offers a reasoning model, an action-model stack and a generated simulator. Rival bets on world models for robot training include NVIDIA Cosmos and World Labs; a pixel world that drifts after minutes is a tool for scenario variety, not yet a replacement for physics engines such as MuJoCo or Isaac Sim.