NVIDIA research veteran Sanja Fidler launches Veeda AI for robot world models

Why it matters

A team that helped shape NVIDIA's world-model research is now trying to own the model layer independently, betting simulated interaction will supply the training scale that robotics still lacks.

Sanja Fidler (@FidlerSanja) launched Veeda AI on Wednesday with longtime collaborators Zan Gojcic (@ZGojcic) and Huan Ling (@HuanLing6), forming a new AI lab to build world models that can train robots inside generated physical environments.

Fidler introduced Veeda AI in a post on X on August 19th. The launch gives a company name and a commercial home to the world-model thesis she developed during eight years at NVIDIA: robots will require realistic, interactive simulations where they can repeatedly act, fail and adapt before their policies are deployed on hardware.

From NVIDIA research to a startup

Fidler announced on July 31st that she was leaving NVIDIA, where she led its Toronto AI research operation and rose to vice president of AI research. She remains an associate professor at the University of Toronto and was a co-founding member of the Vector Institute. Her research has centered on computer vision, 3D scene understanding, simulation and multimodal models, according to her University of Toronto profile.

Veeda's other founders extend that work across reconstruction, generative models and the data systems needed to train them. Gojcic led NVIDIA research in Zurich on neural reconstruction and generative world simulation for physical AI. He earned his PhD in 3D computer vision at ETH Zurich and previously worked as a visiting researcher at Stanford.

Ling completed his PhD at the University of Toronto under Fidler and later became a research manager at NVIDIA. His teams worked on generative world models and high-fidelity foundation models, including contributions to NVIDIA Cosmos, the chipmaker's platform for building and adapting models that simulate physical environments.

The three founders have already spent years publishing and building together. That shared history gives Veeda a technical base spanning video generation, 3D reconstruction, simulation and training infrastructure, rather than a founding team assembled around the current robotics funding cycle.

The world-model bet

Veeda says it is developing multimodal foundation world models that simulate physical reality, with the goal of creating environments where embodied AI agents can learn through interaction. Its public thesis is direct: imitation learning from recorded human behavior will not be sufficient to produce broadly capable robots.

Real-world trial and error is slow, expensive and capable of breaking machines or injuring people. Veeda wants to shift much of that learning into generated environments, which can be duplicated, accelerated and varied by software. The founders describe the intended system as "the Matrix" for physical AI, although Veeda's first task is the less cinematic work of making generated scenes consistent enough for robots to learn useful behavior from them.

The company's hiring operation shows the scale of that task. Veeda is recruiting researchers and engineers for multimodal generation across images, video and 3D, along with distributed training, data curation and machine-learning infrastructure. Its listed locations are Toronto, Zurich, Mountain View and Singapore.

Those roles point to a compute- and data-heavy model company rather than a robotics hardware manufacturer. Veeda is targeting the simulation layer beneath robot policies: the generated worlds, sensor observations and interactive experiences used to train and evaluate machines before deployment.

Competing with the lab they helped build

Veeda enters a field where NVIDIA is already pushing an open model stack. NVIDIA Cosmos combines world generation, physical reasoning and action modeling for robots, autonomous vehicles and other physical AI systems. Its latest family, Cosmos 3, is designed to process and generate combinations of language, images, video, audio and actions.

Fidler's move puts Veeda close to the technical territory she helped establish inside NVIDIA. Ling was a core contributor to Cosmos, while Gojcic's NVIDIA work focused on constructing and simulating 3D environments. Veeda is now turning that research lineage into an independent company whose value will depend on whether its models can improve robot training beyond conventional simulators, recorded demonstrations and synthetic video pipelines.

That is the central technical test. A generated environment can look convincing while producing incorrect physics, inconsistent objects or unrealistic consequences for an agent's actions. Robotics developers need simulations that remain coherent through interaction and transfer into measurable gains on real machines.

Veeda is betting that interactive world models will become the scaling mechanism that internet text provided for language models. Fidler has left one of the best-funded physical AI programs to pursue that bet with two researchers who helped her build it.