DeepMind World Models Train Robots in Imagined Worlds: SIMA Practices Inside Genie 3 Model

▤ Summary

DeepMind is pairing its Genie 3 world model with the SIMA 2 agent to train AI inside generated 3D environments. The setup lets SIMA 2 practice navigation and task execution without new human demonstrations, which DeepMind presents as evidence that agents can improve through self-play in imagined worlds. The company positions this as a way to address the robotics data bottleneck by replacing slow, expensive real-world data collection with scalable simulation. DeepMind also notes clear limits, including short interaction windows, limited action space, and weak long-horizon reasoning. The broader commercial implication is that generated environments could become a major training layer for robotics, autonomy, and agent development.

▦ Linked Companies

DeepMind

$100M to $250M