How Waabi’s Raquel Urtasun Is Building One A.I. Brain to Drive Trucks and Robotaxis

Waabi founder and CEO Raquel Urtasun is taking a “less-is-more” approach to autonomous driving, betting on an A.I. system that can scale without training on mountains of real-world road data. Since 2023, Waabi’s freight trucks, manufactured by Volvo, have been testing on Texas highways with safety drivers, with plans to go fully driverless next year. At the same time, the company is bringing its core technology to passenger cars through a deal with Uber to deploy at least 25,000 robotaxis in the coming years.

This kind of crossover was not possible a few years ago. Early autonomous driving systems relied on engineers manually coding rules for every scenario, while more recent approaches rely on collecting billions of miles of road data to train black-box systems. Urtasun sees both methods as fundamentally unscalable. She founded Waabi in 2021 to pursue a different route. The company trains its system inside a hyper-realistic, mathematically proven simulator before taking it to the road. The result is a single generalizable A.I. “brain” capable of learning from few examples and adapting seamlessly across vehicle types.

Urtasun is a leading scholar in computer vision, a core technology behind autonomous driving. She is a University of Toronto professor and co-founder of Canada’s Vector Institute, a leading machine learning nonprofit. Before founding Waabi, she led Uber’s self-driving research division.

Over the summer, Urtasun spoke with Observer about the technological breakthroughs that changed autonomous driving, Waabi’s expansion into robotaxis, and her plans for a potential IPO.

The following conversation has been edited for length and clarity.

Observer: Human drivers need a specialized license to handle a semi-truck compared to a Sedan. How can a single A.I. system handle both?

A driver with a commercial driver’s license can drive a passenger car, but someone with a regular passenger-car license cannot drive a truck. This implies that driving a truck is more difficult than driving a passenger car. Navigating a massive tractor-trailer, with all the complexities of its physics and weight distribution, is difficult for humans and, in general, difficult for machines. You can understand why bringing that capability to robotaxis is easier than going the other way around.

One of the things that makes this possible is the generalization ability of our technology. It all revolves around whether you can build a brain that can generalize across environments, vehicle types and different sensors.

It means we no longer need to build a different technology stack for every use case, with every program requiring its own staffing and development. For the first time, everything is additive. Every improvement we make to our robotaxis will directly improve our trucking product. For example, robotaxis operate in more urban environments, and the experience we gain there will allow us to expand further into urban areas with our trucks.

Why haven’t traditional self-driving models worked as well as people hoped? Where have previous approaches fallen short?

In 2021, the industry was divided between AV 1.0 and AV 2.0. AV 1.0 consisted of hand-engineered approaches in which you had to manually encode how the system should think. As you can imagine, that doesn’t scale. It’s impossible for a few humans to write rules that cover everything the system needs to do.

Then came the AV 2.0 era. There are many different approaches within this category today. What has become more mainstream, particularly in Level 2+ systems where there is still a human on board, is a black-box architecture. Regardless of whether you’re using cameras, lidar or other sensors, you feed sensory information into the system, and it directly outputs an action without explicit reasoning or thinking. It simply tells you what the vehicle should do.

This type of technology is more similar to large language models. Elon Musk has predicted that Tesla would need 10 billion miles of driving data to achieve self-driving goals. [Tesla hit that goal in March.] That’s the kind of black-box architecture approach that doesn’t scale. It requires so much data that even 10 billion miles is not enough. You need to encounter virtually every type of situation to be able to handle it. It’s also very capital-intensive because you need massive data centers to train and continuously improve the system.

What is Waabi doing differently that allows your A.I. to adapt and reason like a human?

When I started Waabi, we placed a big bet on two new technologies that we believed would deliver the ChatGPT moment for physical A.I.

The first was that, to build something that truly scales and generalizes across vehicle types, environments and situations, you need to take AV 2.0 to the next level. Building a black-box architecture is not enough. You need an end-to-end system capable of reasoning in a verifiable way, so that you can establish and prove the level of safety required to deploy this technology without a human on board.

This was a massive breakthrough. It can learn from very few examples and transfer skills across different platforms and environments. That’s how humans learn. When you have that capability, you need much less data. That translates into a much faster path to deployment, a more scalable approach and much greater capital efficiency.

That gives us two advantages. First, we don’t need as much data. Second, the system can truly generalize to situations it has never seen before, and it does so in a way that allows us to verify and prove safety. That’s extremely important for regulators and for removing the human driver.

If you aren’t collecting billions of miles on real roads, how do you actually train and test the system for dangerous situations?

It was obvious that relying on real-world data alone would never be sufficient. As you’re developing the technology and trying to assess the safety of the system, safety-critical situations are the most important. You want to train the system on those situations and make sure it responds correctly, including in scenarios involving avoidable and unavoidable accidents.

The world is incredibly diverse in terms of everything that can happen. It became clear to us that building a very large fleet and hoping to encounter every possible situation would never be the right approach.

Instead, we need to build a world model, a simulator that is as realistic as the real world. We built a simulator where we can mathematically prove that driving in the simulator is equivalent to driving in the real world.

Waabi initially planned to deploy fully driverless trucks by the end of 2025, but that timeline has been pushed back. What’s holding up the launch?

We are waiting for Volvo to complete the final stages of validation. We’re waiting because of safety. We don’t believe retrofitting a system that has not been fully validated is safe, and we don’t compromise on safety.

Other companies face a lot of pressure from investors to demonstrate progress every quarter. They’re doing things that we believe are unsafe, and we will never do that.

Now that Volvo has made its timeline public, we can share it as well: Q1 2027.

What are you focusing on in the meantime while waiting on hardware validation?

We’ve been very busy working with customers. We have a massive pipeline of Fortune 50 companies that cannot wait for this technology and platform to be fully validated and ready.

Waabi’s solution is the only one in the market that truly meets their needs. If you look at other players in the industry, they’re pursuing a hub-to-hub model. They have hubs next to the highway, and a human driver needs to transport the freight between those hubs and the end customer. The technology operates autonomously only between the hubs.

With our technology, we can go directly to the customer. Since Q1 2025, we’ve been able to operate door-to-door. That’s massive because we can transport freight as it’s transported today without requiring three separate legs, two of which involve human drivers.

Over the next two to three years, our growth is going to be limited by how quickly the OEM can produce trucks.

Given those manufacturing constraints, would you ever consider building your own trucks?

No. OEMs are amazing at what they do. Particularly when it comes to self-driving trucks, you really want safety to be front and center, and we believe OEMs are the only ones that can deliver that today. That’s why we’re waiting for them rather than focusing on demonstrating driverless milestones.

Back in Q3 2025, we demonstrated driverless operations on a closed course. We built trucks with redundant systems in-house, but we made the decision that, when it comes to operating on public roads, the platform should come from the OEM.

With this growth outlook, is an IPO on the horizon?

I get asked about it quite a bit these days. We’re definitely in no rush. At the same time, given how fast we’re growing, we’ll see if and when the time is right for us.

We just raised over a billion dollars in Q1, and given how capital-efficient we are, that gives us many years of runway. We have the luxury of deciding when we want to go public. We can pursue a traditional IPO rather than going public through a SPAC.

We don’t have specific plans as of today, but you can imagine what the IPO opportunity could look like once we have predictable, significant revenue across robotaxis and trucks.

But no SPAC.

Definitely no SPAC. Companies often pursue SPACs when they have difficulties raising capital in the private markets. Going public through a SPAC is complex and can put a company on a difficult path.

You’re also a full professor at the University of Toronto. What does your role there involve, and how do you balance it with running Waabi?

My role at the university right now is to advise the university president, Melanie Woodin. That’s definitely not a full-time job, but it’s an important one.

My experience as an academic, in big tech during my Uber days and now as an entrepreneur can help the university understand how it should position itself in this new world of A.I.

The goal is to ensure that the University of Toronto continues to be a leading institution, both in research and in creating more entrepreneurship opportunities for students.

I also have Ph.D. students. That’s another connection to the university, but those students are embedded full-time at the company.

This is a very exciting new model of education. For physical A.I., it’s the only way you can really do cutting-edge research. You need to be fully embedded within a cutting-edge company.