Environment simulation instead of passive video. The same technology is the trainer where physical skills are cheaper and safer to learn than on real hardware.
About this eventReal-time interactive worlds with scene consistency lasting several minutes.
Genie 3 generates a controllable environment from text and supports real-time interaction. It is not an accurate physics simulator: action space, long-term consistency and interaction between independent agents remain limited, but it can already serve as a research environment for training and evaluating behaviour.
Source: Google DeepMindAbout this eventPlayable generated worlds opened to some Google AI Ultra subscribers in the US.
About this eventThe standalone app closed and the API ends on 24 Sep 2026 — video generation folds into general models.
About this eventGoogle I/O 2026 unveiled a model that creates video from any input.
About this eventAn open physical-AI omnimodel combined vision, world simulation and action generation in one architecture.
About this eventObjects fall and bounce plausibly, but the model guesses physics rather than computing it. On long scenes it shows.
About this eventA skill learned in simulation must work on real hardware. The sim-to-real gap is the main obstacle.
About this eventA fully traversable environment training not just robots but agents — from driving to running a factory.
Primary material behind this dossier: papers, lab publications and official reports.