World models, the AI systems that learn how the physical world behaves and then predict what happens next, crossed from research demo to practical infrastructure in 2026. Google DeepMind's Genie 4, announced on October 27, is the most capable example so far: a real-time simulation that renders at 720p, responds to keyboard and text input, generates its own audio, and moves people in a lifelike way. The implications for robot and agent training are significant, but so are the limits. This piece explains what a world model is, what Genie 4 adds over Genie 3, why the technology matters for training, and where it still falls short.
A world model in plain terms
A traditional robotics simulator is a calculator with rules written by engineers. Nvidia's Isaac Sim, for example, computes mass, friction, and contact from hard-coded equations. A world model takes a different approach: it learns from data, often video, and then predicts how a scene will evolve. Amey Deosthali, head of robotics at AMD, describes it as a "predictive representation of the physical world that allows an AI system to reason about what is likely to happen." The distinction matters because a learned model can generalize across situations its programmers never explicitly modeled, and it can be trained on the enormous amount of video already on the internet.
The most ambitious versions, called world action models, go further and output the robot's movements as well as the predicted scene. Runway's Praxis-1, released with open weights on November 11, 2026, predicts the next moments of a video and the matching arm and gripper movements in one system. Other labs shipped similar systems through the year. Genie 4 sits in the adjacent category: it generates the world, and a separate Gemini model handles the reasoning and planning that act inside it.
What Genie 4 adds over Genie 3
Genie 3, previewed in August 2025, established the basic idea: a general-purpose world model that could generate interactive environments at 720p and 24 frames per second, holding consistency for a few minutes, with real-time navigation and promptable events. It was a research milestone, but the sessions were short and the visuals, while impressive, were clearly synthetic.
Genie 4 pushes four capabilities at once. The first is real-time steering: you can change the weather, move the camera, or introduce a person or object with natural language, and Gemini adapts the request to match the scene while keeping the world coherent between inputs. The second is physics and human movement. Genie 4 simulates aerodynamics, collisions, and other physical scenarios, and its characters move naturally, respond to their environment, and can be directed interactively. The third is native audio. Genie 4 is Google's first real-time world model to generate sound as part of the simulation, from wind in trees and waves on a beach to interactive dialogue that adapts on the fly. The fourth is complex events: Gemini 3.8 Flash acts as a guiding agent, watching the rendered frames, decomposing a goal into steps, and verifying each one, which makes multi-step tasks reliable rather than improvised.
Why this matters for training robots and agents
The core problem in robotics is variation. A robot that picks up a cup needs to handle differences in weight, texture, and fill level, and testing every variant on real hardware is expensive and sometimes dangerous. A world model can generate those variants cheaply and safely. As Jonathan Hurst of Agility Robotics notes, testing on real robots is costly, and robots "can fall, hit, and hurt people."
Google has already shown the pipeline. Gemini 4 Argon watched a video of a person playing cornhole, mapped the environment, calculated the joint movements, built the simulation code, and transferred the result to Apptronik's Apollo humanoid, which sank seven consecutive shots in a real trial. Genie 4 generalizes that idea into a reusable training ground. DeepMind describes safe, infinitely variable simulations as a crucial step toward AGI, and the company has experimented with Argon and Genie 4 together to simulate Apollo playing checkers, with Argon feeding step-by-step captions into the live simulation.
Independent evidence suggests the approach is maturing. Google has also experimented with predicting how robot policies will play out and checking those predictions against real trials. Benchmarks such as RoboLab-120, a set of tabletop tasks built in Isaac Sim, have become a standard way to grade these policies in a controlled environment.
The honest limits
DeepMind is unusually direct about Genie 4's weaknesses. The model maintains coherent environments over short sessions, but visual details and object permanence can drift over time, especially in longer runs, after many prompt edits, or while navigating. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency are all listed as active research areas. There is also no benchmark table in the announcement and, as of late October 2026, no independent third-party evaluation, so the realism claims rest on Google's own demonstrations.
The deeper issue is the sim-to-real gap. A simulated score is only as good as the simulation, and small prediction errors compound over time. Contact-rich manipulation, such as handling a full cup of water, remains difficult. Hurst's warning is worth remembering: "simulations are doomed to succeed." You can show what you want in simulation and still fail in the real world. Deosthali adds that appearance alone cannot reveal every physical property needed to handle an object. Practical teams therefore treat world models as a way to generate training variety and triage policies, not as a replacement for real hardware validation. Nvidia's grading approach, which tests policies in a conventional simulator where object positions are known exactly, remains the safer exam.
Access is another constraint. Genie 4 is a research preview for trusted testers, with a waitlist for Project Genie and early access through the Gemini API and Google AI Studio. The API generates 720p video at 24 frames per second with synchronized audio, priced at $17.50 per million output tokens including audio, which works out to roughly $0.10 per second. That is affordable for experimentation but not free, and every output carries an imperceptible SynthID watermark.
What to watch next
DeepMind's stated next step is hybrid world models, which combine the structured reasoning of language and code models with pixel-level simulation. Gemini would map the structure and plan the timeline while Genie renders how the world looks, sounds, and responds. If that works, it addresses the exact failure mode DeepMind admits to: long-session coherence. The near-term signs to watch are independent benchmarks, broader API access, and whether robotics teams can show policy improvements that survive contact with real hardware.
If you want a clear-eyed view of how world models and other frontier AI systems affect real business decisions, Crowe Endeavors follows the releases, the pricing, and the practical trade-offs.
Sources
"Genie 4: A new generation of world simulation" (Google DeepMind, October 27, 2026) — https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/
"Genie 4 overview" (Google (YouTube), October 2026) — https://www.youtube.com/watch?v=sFfPkUJd_Jk