Google DeepMind’s Genie 4 lets users change a simulated environment while exploring it, using natural language prompts. They can alter the weather, move the camera or add people and objects, with audio generated in real time.
The model extends scene consistency from about 1 minute in Genie 3 to about 2 minutes. Gemini 3.8 Flash helps guide the simulation by breaking down goals and checking output frames.
Google is pitching the technology for uses including robot training scenarios, and has demonstrated a robotics experiment involving simulated tasks. However, objects and details can drift during longer sessions, while audio accuracy and readable text remain unresolved.
Access through Project Genie is due to expand to Google AI Ultra subscribers in the coming weeks, with early access also planned for Gemini API partners. For companies exploring simulation tools, this is an opportunity to test a research model rather than evidence of a dependable training environment.