World foundation models — AI systems that simulate how the physical world actually behaves, rather than just predicting the next word — have become the bottleneck determining how fast humanoid robots can learn, and NVIDIA's Cosmos and Google DeepMind's Genie 3 represent the two most consequential approaches to solving it, built around genuinely different priorities: Cosmos optimizes for strict physical accuracy robots can actually train on, while Genie 3 optimizes for generating novel, never-seen environments in real time. This is a direct follow-up to our comparison of GR00T, Gemini Robotics, and Tesla Dojo — this piece goes one layer deeper, into the simulation engines actually generating the training data those robot brains learn from.
Why World Models Matter More Than the Robot Body
The scaling law NVIDIA identified with GR00T N1.7 — more training data reliably produces better robot dexterity — only matters if you can generate that data affordably. Real-world teleoperation data collection is slow and expensive; a human operator can only generate so many hours of footage. World foundation models solve that bottleneck by generating synthetic training environments and trajectories computationally, which is exactly how NVIDIA's GR00T-Dreams pipeline compressed synthetic data generation from three months down to roughly 36 hours, discussed in the previous entry in this series. The two platforms compared here represent the current frontier of that synthetic-data infrastructure.
NVIDIA Cosmos 3: Built for Physical Accuracy at Industrial Scale
NVIDIA's Cosmos 3, an open world foundation model, was trained on roughly 20 trillion tokens of multimodal data spanning nearly a billion images, 400 million real and synthetic videos, and audio and action data captured from both humans and robots. It's omnimodal by design — jointly modeling images, video, audio, and actions within one unified framework, rather than treating each type of data as a separate system. Since launch, Cosmos has been downloaded more than 2 million times, and NVIDIA has organized a Cosmos Coalition with partners including Agile Robots, Black Forest Labs, Runway, and Skild AI to advance the platform collaboratively. Its defining technical priority is strict physical consistency: the model is built to maintain accurate physics — how objects fall, collide, and interact over time — specifically because industrial and robotics applications can't tolerate a simulation that looks realistic but behaves in physically impossible ways. Companies like FieldAI and Skild AI already use Cosmos specifically to generate training data at a scale real-world collection can't match.
Google DeepMind Genie 3: Built for Novel, Interactive Worlds
Genie 3 takes a different starting point. Released in research preview, it generates photorealistic, interactive 3D environments from text or images at 24 frames per second in real time, supporting persistent worlds with object permanence and what DeepMind describes as emergent physics — behavior that arises naturally from the model rather than being explicitly programmed. Genie 3 powers Project Genie for Google AI Ultra subscribers and is positioned as a tool for both agent training and creative environment generation. Where Cosmos prioritizes physical fidelity for industrial reliability, Genie 3's strength is generating genuinely novel environments from a simple text description — a more creative, exploratory capability that trades some of Cosmos's strict physical grounding for flexibility in generating scenarios that have never existed in any training data.
The Core Tradeoff: Physical Accuracy vs Creative Generation
Industry coverage has converged on a clean way to describe the split: Genie 3 excels at generating novel environments from text, while Cosmos maintains strict physical consistency for industrial applications. That's a genuinely different design philosophy, not just a feature gap — Cosmos is built to be trustworthy enough that a robot trained on its synthetic data behaves correctly when it encounters the real world, while Genie 3 is built to be flexible enough to imagine environments and scenarios a robot (or an AI agent more broadly) might never have encountered in any dataset. Complementary physics engines like Genesis and MuJoCo occupy a third position entirely: Genesis reportedly reaches 43 million frames per second on a simple robotic-arm scene using an RTX 4090, far faster than either world model, but produces no photorealistic visuals — teams increasingly combine fast physics-only simulators like Genesis for rapid policy search with a world model like Cosmos-Transfer for realistic visual domain adaptation before final deployment.
Other Contenders Worth Knowing
Cosmos and Genie 3 aren't the whole field. Fei-Fei Li's World Labs launched Marble, making world model generation commercially available with pricing ranging from free to $95 a month, and was reportedly in talks for a valuation around $5 billion. Yann LeCun left Meta after 12 years specifically to launch AMI Labs, raising roughly €500 million at a €3 billion valuation to build systems grounded in physical understanding using JEPA-style architectures rather than the diffusion or autoregressive approaches Cosmos and Genie 3 use. Neither Marble nor AMI Labs has the direct, established tie to humanoid robot training pipelines that Cosmos has built through NVIDIA's broader Isaac ecosystem, but both are credible signals that world models are becoming a genuinely competitive category rather than a two-company race.
Which One Actually Matters for Robotics?
For teams building humanoid robots specifically, Cosmos currently has the more direct, proven connection to production robot training — it's the backbone behind GR00T-Dreams' synthetic data pipeline, already powering real deployed robots across NVIDIA's broad hardware partner ecosystem covered in the previous entry in this series. Genie 3's robotics applications remain earlier-stage and more research-preview in framing, with its clearest current use case centered on interactive environment generation and agent training more broadly, rather than a dedicated robotics production pipeline on the scale of Cosmos-Transfer. That could shift as Genie 3 matures out of research preview, but as of mid-2026, Cosmos is the world model with the clearer, more direct line to robots actually shipping.
World Foundation Models at a Glance
| Model | Core strength | Training scale | Robotics maturity |
|---|---|---|---|
| NVIDIA Cosmos 3 | Strict physical consistency for industrial reliability | ~20 trillion tokens; ~1B images, 400M videos | Production — powers GR00T-Dreams synthetic data pipeline |
| Google DeepMind Genie 3 | Novel, interactive 3D environment generation from text | Not fully disclosed; real-time 24fps generation | Research preview — broader agent training focus |
| World Labs Marble | Commercially accessible world model generation | Not fully disclosed | Early; creative and spatial-intelligence framing |
| AMI Labs (LeCun) | JEPA-based physical understanding, distinct architecture | Not fully disclosed; €500M raised for development | Early-stage, architecture still maturing |
Frequently Asked Questions
What is a world foundation model, and how is it different from a regular AI model?
A world foundation model learns how the physical world behaves over time, how objects move, collide, and interact, rather than learning from text alone. This lets it generate synthetic training environments and predict physical outcomes, which is essential for training robots without needing constant real-world data collection.
What's the main difference between NVIDIA Cosmos and Google DeepMind Genie 3?
Cosmos prioritizes strict physical accuracy, making it reliable for industrial and robotics training data. Genie 3 prioritizes generating novel, interactive environments from text in real time, trading some physical grounding for creative flexibility.
Which world model is actually used to train real robots today?
NVIDIA Cosmos has the more established, production connection to robotics, powering the GR00T-Dreams synthetic data pipeline already used across NVIDIA's humanoid robot partner ecosystem. Genie 3 remains in research preview with a broader focus on agent training and interactive environments.
Are there other companies building world foundation models besides NVIDIA and Google?
Yes. Fei-Fei Li's World Labs launched Marble as a commercially available world model tool, and Yann LeCun left Meta to found AMI Labs, raising roughly €500 million to pursue a JEPA-based approach to physical world understanding.