top of page

Beyond the Chatbot Sandbox: Sanja Fidler’s Veeda AI Wants to Build a Training Ground for Robots.

23 hours ago
4 min read

An emerging company, an ambitious simulation strategy, and the question that will determine whether physical AI delivers: does virtual practice improve real-world performance?
An emerging company, an ambitious simulation strategy, and the question that will determine whether physical AI delivers: does virtual practice improve real-world performance?

There’s a moment in every technology cycle when the conversation shifts from what a machine can generate to what it can reliably do.

In robotics, that question carries physical consequences. A convincing answer on a screen is one thing. A robot moving beside a person is another.

That is what makes Veeda AI worth watching beyond its August debut.

The Toronto-headquartered company is led by former NVIDIA vice president of AI research Sanja Fidler, alongside co-founders Zan Gojcic and Huan Ling. Its financing was announced on August 19, 2026, with backing from Radical Ventures and Khosla Ventures, according to Osler, the law firm that advised Veeda on the transaction.

BetaKit reported US$90 million in funding. That figure should be treated as a reported amount: Osler’s transaction announcement confirms the financing and investors but does not disclose its value. [1, 2]

The larger story is the proposition behind the investment: robots may need far better places to practice before we ask them to operate around us.


Meet the woman building a virtual classroom for physical AI



Sanja Fidler, ex-NVIDIA who raised $90M to build world models for Physical AI at her VEEDA AI startup.
Sanja Fidler, ex-NVIDIA who raised $90M to build world models for Physical AI at her VEEDA AI startup.

Fidler previously led NVIDIA’s Spatial Intelligence Lab. Radical Ventures describes her research background as spanning 3D scene synthesis and generative AI. The investor also identifies Gojcic as a former NVIDIA research director and Ling as a former NVIDIA research manager focused on generative AI.

At Veeda, Fidler is CEO, Gojcic is CTO, and Ling is chief scientist, according to launch coverage. Fidler also has longstanding academic ties to the University of Toronto and was a founding member of the Vector Institute.

The team’s ambition begins with the environment in which a machine learns. Veeda’s website describes a company developing multimodal foundation world models for interactive robot learning, with listed locations in Toronto, Zurich, the Bay Area, and Singapore.

This is a research and infrastructure ambition. It should not be confused with proof that Veeda has already solved general-purpose robotics.


What is a world model?

A useful way to think about a world model is as a learned predictor of how an environment changes. In an interactive setting, an agent supplies an action and the model predicts a resulting observation or state. NVIDIA’s current simulation work, for example, generates visual observations in response to camera or robot actions.

Imagine a robot reaching toward a cup.

Recognizing the cup is a perception task. Predicting how the scene might change when the gripper moves is a different challenge. A useful predictive model could help a system evaluate possible actions before attempting them physically.

But the distinction matters: predicting plausible outcomes does not establish humanlike understanding, accurate physics, or safe behavior. Those require evidence.

A beautiful simulation can still teach the wrong lesson.


Why practice in simulation?

Veeda argues that physical trial and error is too costly, slow, and potentially unsafe to serve as the sole route to learning at scale. Its proposed alternative is simulated reality where embodied agents can repeatedly interact with generated environments.

The appeal is straightforward. A failed virtual attempt need not damage a machine, interrupt a facility, or put someone in its path. Developers could vary conditions and expose systems to a broader range of situations before physical testing.

That makes simulation a potentially powerful complement to demonstrations, conventional simulators, and real-world data. It does not make those other tools obsolete.

For a warehouse operator or manufacturer, the relevant question is practical: does this training approach produce a more reliable system at a reasonable cost?


From Pac-Man to driving simulation

Fidler’s earlier research helps explain the trajectory.

She co-authored GameGAN, presented in 2020, which learned to emulate a game environment from gameplay data. NVIDIA demonstrated a generated version of Pac-Man that responded to an agent’s actions without using the original game engine.

She also co-authored DriveGAN, presented at CVPR 2021. Its researchers trained a controllable neural simulator on multiple datasets, including 160 hours of real-world driving data. The project demonstrated controls for steering and speed, alongside changes to generated scene components.

These were research steps toward learned interactive environments. They were not demonstrations that simulation alone could certify a safe autonomous vehicle.

That boundary remains essential as the category attracts more attention.


What has changed since August?

The strongest update is broader technical progress in interactive simulation.

In a November 2, 2026 technical post, NVIDIA described Cosmos Framework workflows and released checkpoints for generating simulation observations across autonomous vehicle, robotics, and surgical scenarios. Its accompanying research page demonstrates action-conditioned visual generation, spatially controlled scene generation, and multi-view driving simulation.

These are NVIDIA developments, not Veeda product announcements. They provide useful context: the environment-generation layer Veeda is targeting is an active area of development with substantial existing competition.


The reality check: a convincing world can still be wrong

For this category, the editorial test should extend beyond impressive video.

Does an object remain consistent through a long interaction? Does a model respond correctly when a robot makes an unfamiliar move? Does training transfer to physical hardware? What happens when sensors are noisy or the environment changes?

Those are the questions developers and buyers should ask before treating a generated world as a dependable training environment.

Useful evidence would include comparisons with existing approaches, clearly described hardware tests, failure rates, and the cost of achieving a measurable improvement. The goal should be better performance outside the simulation.


Why Veeda still belongs on the watch list

Fidler’s move is compelling because it combines research leadership with a concrete commercial question: can simulated experience become a more effective way to develop physical AI?

Her leadership also deserves coverage on its technical and business merits. A woman founding an ambitious AI infrastructure company is a meaningful story; her research record gives it substance beyond representation alone.

The opportunity here is to follow the evidence as closely as the ambition. Veeda’s financing and team make it worth watching. What comes next will determine how much of that promise becomes useful infrastructure.

The future of AI may include machines that speak more fluently. The harder milestone is a machine that acts more reliably.

And reliability will have to be demonstrated in the world we actually inhabit.

Subscribe to Receive Our Latest Tech News

About Us

Tech Deck Magazine is where AI, innovation, and culture converge. Through original editorial, thought leadership, and exclusive coverage of the people, ideas, and industries shaping what's next, we connect founders, executives, innovators, and changemakers across technology, business, and influence. Our mission is to inform, inspire, and cultivate a thriving ecosystem where innovation and opportunity flourish.

bottom of page