Google has unveiled Genie 3, a new “world model” designed to simulate realistic environments for training AI robots and autonomous vehicles. Developed by Google DeepMind, the model creates detailed virtual settings like warehouses where AI agents can learn to perform tasks autonomously.
World models such as Genie 3 are considered crucial for advancing artificial general intelligence (AGI), which aims to enable AI systems to handle diverse human-level tasks across different domains, rather than specialized functions like playing chess or language translation.
DeepMind highlighted that world models will play an essential role as AI agents assume more active roles. Genie 3 generates immersive scenarios directly from text prompts, allowing users to quickly adjust environments—for example, adding a herd of deer to a ski slope simulation.
Though Google demonstrated Genie 3’s skiing and warehouse simulations, the model is not yet publicly available and has notable limitations. The company has not announced a release date. Google also introduced Sima, a virtual agent capable of task execution within game environments, which remains inaccessible to the public.
Experts emphasize the significance of world models for robotics. Professor Subramanian Ramamoorthy of the University of Edinburgh noted that anticipating outcomes in physical settings is essential for flexible robot decision-making. Andrew Rogoyski of the University of Surrey added that virtual embodiment through world models could enhance AI capabilities beyond data-driven large language models.
The announcement coincides with intensifying AI competition, including OpenAI’s recent preview of GPT-5. Google envisions Genie 3 aiding both AI training and human experiences by enabling simulation of activities like skiing or exploring natural landscapes.