Google DeepMind Releases Research Preview of World Model Genie 4

Download the app

+1 Google Releases New-Generation World Model Genie 4, Strengthening Physical AI Technology Layout

The Paper reporter Yu Yan 2026-10-28 11:31 Source: The Paper ∙ Sci-Tech 101 > Listen to full text Font size

On October 27 local time, Google's DeepMind released a research preview of its new-generation world model Genie 4. The model further improves real-time interactive world simulation capabilities while exploring applications in robot training and the field of Physical AI.

Compared with the previous-generation model Genie 3, Genie 4 not only improves visual realism, physical simulation and the naturalness of character movements, but also extends the duration of simulations that maintain scene consistency to about two minutes, twice that of the previous-generation model.

According to the introduction, Genie 4 can generate the most realistic real-time interactive dynamic virtual worlds to date, with progress in simulating physical phenomena, character movements and scene interaction. Users can use natural language instructions to change the weather, adjust the viewing angle, or add characters and objects in real time during simulation. The model also has native audio generation capabilities and can synchronously generate background music, ambient sounds and interactive dialogue.

Notably, Google also demonstrated an experiment in which Genie 4 worked with the large model Gemini 4 Argon to simulate a humanoid robot performing tasks.

In the demonstration, Gemini 4 Argon was responsible for breaking user instructions into multiple steps, while Genie 4 generated, in real time according to these instructions, simulation footage of the robot performing the task, showing the Apollo humanoid robot from the U.S. robotics company Apptronik playing checkers.

This experiment also demonstrates Google's technology layout in the field of Physical AI: an attempt to combine the task planning and reasoning capabilities of large models with the environment simulation capabilities of world models, enabling AI to simulate different tasks and their execution processes in virtual worlds.

Comparison image. All images above are from: Google DeepMind official website

Tejas Kulkarni, Google's director of research, said at a media briefing before the release that more and more "hybrid world models" will emerge in the future, in which Gemini generates code and text, and Genie receives this information in real time and simulates the real world at different levels of abstraction.

World models are becoming an important technology direction for tech companies. Potential applications cover not only game development, interactive education and film and television production, but also robot training, autonomous driving simulation and other fields. By generating interactive virtual environments, world models are expected to help robots and AI agents conduct task simulation and testing before entering real environments.

Industry competition and integration are also accelerating. In September this year, AI video generation company Runway released its second world model, GWM Worlds 2, which supports dynamic generation of interactive scenes based on user input and has audio generation capabilities.

On the capital market side, on September 28, AMD announced that it would acquire World Labs, founded by Chinese-American AI scientist Fei-Fei Li, in an all-stock transaction worth about $8.2 billion, strengthening its technology layout in world models, spatial intelligence and Physical AI. Unlike Google's Genie 4, which focuses on real-time dynamic simulation, World Labs' previously launched Marble mainly focuses on generating three-dimensional worlds with spatial consistency that can be explored and edited over the long term.

In addition, according to The Wall Street Journal, Amazon is in talks to acquire world model startup Decart for about $7 billion. On October 22, U.S. electric vehicle manufacturer Rivian announced the acquisition of world model startup Third Dimension and plans to apply the related simulation technology to the development and testing of autonomous driving systems.

However, world models still face technical challenges before large-scale practical application. DeepMind engineer Diego Rivas said at a media briefing that Genie 4 is still at an early stage, and the complete product application form is not yet clear. As simulation time lengthens, details and objects in the scene may drift, and issues such as audio accuracy and the clarity of text in images remain to be resolved.

Google plans to open it to Google AI Ultra subscribers through Project Genie in the coming weeks, and gradually provide the related capabilities to early access partners of the Gemini API and Google AI Studio. Responsible editor: Huan Yanhong Image editor: Shen Ke The Paper news tip: 021-962866 The Paper, unauthorized reproduction is prohibited +1 Favorite I want to report #Google #WorldModel #PhysicalAI View more