Google DeepMind is advancing its world simulation model with Genie 4, capable of generating more realistic environments, producing their soundtracks, and modifying scenes on demand. Still in the research phase, this technology targets robotics, video games, and audiovisual previsualization.
- Google DeepMind’s Genie 4 generates interactive environments, soundtracks, and scene changes on demand.
- Users can modify a running simulation with natural language instructions.
- The model keeps scenes coherent for about two minutes, twice as long as Genie 3.
- Still experimental, Genie 4 is first offered to limited testers and targets robotics and video games.
After generating text, images, and videos, AI players are focusing on another capability: simulating environments where users can move and act. Unveiled on October 27, 2026, Genie 4 is part of this evolution. Developed by Google DeepMind, this world simulation model (world model, meaning a system that reconstructs a virtual environment and its workings) allows for the creation of interactive scenes whose appearance and behavior evolve based on user input.
Environments controlled by natural language
Genie 4 first stands out by allowing modifications to a simulation while it’s running. An instruction in natural language, meaning phrased in an ordinary sentence, can be used to change the weather, move the camera, or introduce a new object or character. The system must then adapt the scene without breaking its continuity.
The model also generates sounds associated with the environment, whether ambient noises, music, or interactive dialogues. Google DeepMind also indicates integrating capabilities from its Gemini family to improve scene coherence and the management of multi-step events. This combination aims to bring the visual representation closer to the reasoning needed to evolve an environment.
Genie 4 can maintain a coherent scene for about two minutes, according to TechCrunch, which is twice as long as its predecessor Genie 3. However, this duration remains limited for complex or extended simulations.
From digital creation to robotics
The envisioned applications go beyond entertainment. In video games, such a model could accelerate the creation of explorable prototypes before development in a traditional engine. In audiovisual production, according to testers, it could facilitate scene previsualization by allowing quick testing of framing, lighting, or movements.
Robotics is another potential outlet. A simulated environment could be used to examine the behavior of an agent—meaning a software or robotic system that acts autonomously—in various situations before confronting it with real-world conditions. Google DeepMind notably describes experiments combining Genie with Gemini models capable of breaking down a task into steps and then monitoring its execution in the simulation.
These prospects, however, remain to be validated in operational conditions. A visually convincing scene does not guarantee that the simulated physical laws, interactions, or reactions are reliable enough to train an autonomous system.
A still experimental technology
Google acknowledges several limitations: objects and certain details can drift over time, complex physical interactions remain improvable, and texts integrated into scenes are not always legible. Sound accuracy and long-term coherence are also part of the research focus.
The model is initially offered to a limited group of testers. Access is then set to expand to Google AI Ultra subscribers via Project Genie, as well as to certain partners of the Gemini API and Google AI Studio. No widespread deployment is announced at this stage.
According to these initial details, the appeal of Genie 4 will depend less on its visual realism in short sequences than on its ability to maintain coherent, predictable, and controllable environments over time.
