NVIDIA's Cosmos 3 Bets That the Path to Useful Robots Runs Through One Unified Model
NVIDIA released Cosmos 3, a foundation model for physical AI that fuses language, vision, audio, and action into a single architecture — and bundles a lightweight version meant to run on the robot itself.
NVIDIA has shipped Cosmos 3, and the most important detail is structural: rather than stitching together separate perception, reasoning, and control models, the company is positioning a single foundation model as the brain for physical AI. According to @junfanzhu98, Cosmos 3 is built on a Mixture-of-Transformers that unifies language, image, video, audio, and action while jointly modeling world simulation. The pitch is that a robot policy shouldn't have to bolt a world model onto a language model onto a controller — it should reason about all of them in one place.
The architecture being discussed is a dual-tower design, a framing echoed by @MinusWells, who described the release in characteristically breathless terms but landed on the substantive point: this is NVIDIA trying to make world modeling a first-class capability rather than a research curiosity. The dual-tower approach typically separates a generative simulation pathway from an action-prediction pathway, letting the model both imagine future states and decide what to do about them. If that holds up under scrutiny, it addresses one of the persistent failure modes in robotics: policies that can perceive but can't anticipate.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.