Veeda AI Raises $90M to Make Robot Training Less Physical

Veeda AI has emerged with a $90 million seed round and a narrow proposition for physical AI: the limiting resource for robot learning is not only compute, but the inability to safely and cheaply repeat experience in the real world. The Toronto startup is building multimodal world models intended to give embodied AI agents scalable simulated environments for that work.

Veeda Innovation Inc., operating as Veeda AI, announced the financing on August 19, roughly three months after its founding. Khosla Ventures and Radical Ventures co-led the round. One account described the financing as $90 million; another put it at more than $90 million, a small difference that leaves the disclosed scale of the seed investment clear.

The company’s case turns on a contrast between learning on physical hardware and learning inside a modeled environment. Sanja Fidler argues that real-world trial and error is hard to scale because hardware cannot be parallelized in the way compute can; mistakes can also be dangerous. Veeda’s alternative is a simulated setting where robot systems can repeatedly attempt tasks, fail and adjust without relying solely on fleets of physical machines.

World models, as Veeda describes them, are generative systems that learn from sensor and physical-world data to model simulated reality with the quality, variety and physical realism robots require. The intended result is not merely a visual simulation: an embodied agent would operate through a virtual version of its body across different objects, tasks and conditions.

The founders’ relevant starting point

• Sanja Fidler, Zan Gojcic and Huan Ling founded Veeda after working as Nvidia AI researchers.
• Fidler joined Nvidia in 2018 to help establish its Toronto research unit, which later became the company’s Spatial Intelligence Lab.
• Fidler and Ling later helped develop Nvidia’s first world models for physical AI developers.

Veeda’s founders are presenting simulated reality as a potential infrastructure layer for robotics, rather than a single robot application. That positioning makes the financing a bet on a broad training-and-evaluation platform. It also places the company’s central test ahead of it: whether models trained on physical-world data can create environments useful enough for robot learning and evaluation at scale.