TORONTO — Raquel Urtasun wants to fix what she says are shortcomings in the autonomous-vehicle industry. The leading AI researcher and founder and CEO of Waabi has been a leader in the self-driving car industry, dating back to her role as chief scientist and head of R&D at Uber’s Advanced Technologies Group in Toronto. Now she’s out on her own, taking the Toronto-based AV startup out of stealth mode last year and recently unveiling its first major product, a self-driving-truck simulator called Waabi World.

In a conversation with The Logic, Urtasun revealed her thoughts on what regulators are missing about AVs, why self-driving cars can be safe but may never be perfect, and where Canadian talent outshines Silicon Valley.

This interview has been edited and condensed for clarity.

Now that Waabi World is out in the world, you’ve suggested this type of simulation might be useful for regulators. What would that potentially look like?

Regulators today really do not have the appropriate tools to make a very informed decision, in terms of understanding what the technology is for the different competitors. Instead, they have to rely a lot on what the competitors say, versus a really objective way to assess progress and performance. One of the key pieces to showcase safety is simulation, but the current technology out there in the industry falls short. They are also tailored to very specific systems.

It’s not something that we have a contract with the government or anything like that, but I’m happy to put it out in the open that I think this is a key piece for ensuring safety.

You said some of the existing AV simulation technology falls short. How so? 

Many fronts. What you will see in the industry is typically the companies have multiple simulators, not just one canonical big simulator. It’s not able to really have the ability to replace real-world driving, because it doesn’t scale. You don’t see how your reactions, how your driving affects everything. You need the whole thing in order to really mimic how the world works. 

The Waabi World announcement explained the technology as a sort of school for AVs, or a video game where the computer can anticipate your moves and make harder levels every time. So when is the vehicle ready to graduate? What’s the final boss of this video game?

We can run a parallel world, the real and the virtual, and then see whether our system does the same thing. We have the metrics of how realistic it is. Then, you can make a very informed decision that the system is ready for on-road testing and then later on, for removing the driver.  

When you deploy or test your technology in the real world, you’re not going to deploy everywhere at once. Depending on the type of roads, when you’re driving in the city vs. not, certain things will never happen and certain things will happen. You can have all those things in your simulator, so you’re matching the statistics of the real world to the statistics of the simulator.  Based on that performance, you have a notion of, “What is the risk of putting this technology in the real world?”

In a recent webinar, you said that you envisioned starting testing in U.S. markets because self-driving trucks can’t circulate on Canadian roads—but that you’d want to start in Ontario, if you could. What would the government need to do to allow you to test here?

This is a question to ask the government, as well. But they have, for example, a pilot program, or they used to have one, that is for convoys. Convoys are trucks that follow each other, and the first one is human-driven. This is not a type of technology that really is going to deploy. Many years ago, people thought that this makes sense at some level, but it’s not the case. So we will need a pilot program for self-driving trucks on Canadian highways or Ontario highways, but right now it’s not possible to test or operate on Ontario roads.

I’m curious what common fears that people have about AV testing are warranted fears? Which ones are the same kind of thing that would happen if a human was driving? 

I’m obviously not going to comment on [other companies’] individual episodes. But right now, when you want to test self-driving technology in the real world, it’s very important that you have a safety driver, you have a backup system, and it’s very important that you have the proper safety assessments prior to that moment. Any company in this domain should be doing that. 

This [AV] technology has the ability to save many, many of those lives [lost each year]. The question is, does it need to go to zero before deployment? There is maybe a situation where a crash is the only option, regardless of whether you’re the best human driver in the world or a self-driving vehicle. 

People typically expect perfection from machines, but sometimes perfection is not possible. It doesn’t mean they’re not useful or safer. It’s not like I can say, “That accident didn’t happen, and therefore it’s good,” because that accident is not there in the first place. 

If you look at other ways of automation that have happened, whether it is trains or planes, people are always at the beginning a bit skeptical. Then you forget.

The other thing that oftentimes, maybe, confuses people is that the errors of machines are different than the errors of humans.

And sometimes these errors by machines might look dummy to a human because a human can do it. But maybe what they’re not thinking is that something that they cannot do, maybe the machines are doing really well.