DeepMind World Models Train Robots in Imagined Worlds: SIMA Practices Inside Genie 3 Model
AI-summarised brief · reviewed before publication
Google DeepMind has combined two artificial intelligence systems, Genie 3 and SIMA 2, to create a training engine that allows robots to learn in imagined worlds. Genie 3 generates navigable 3D environments from text prompts, while SIMA 2 learns to act within these environments. In November 2025, DeepMind demonstrated SIMA 2 improving its tasks without new human examples, showcasing AI agents can be trained inside invented worlds. This arrangement, called the "Infinite Training Loop," addresses a long-standing issue in general-purpose robots. Genie 3 generates worlds at 24 frames per second, allowing for real-time navigation.
💡 Why It Matters
- · The "Infinite Training Loop" revolutionizes robot training by eliminating the need for physical environments and human-supplied examples.
- · DeepMind's approach enables unlimited curriculum possibilities, paving the way for significant advancements in artificial general intelligence.