If there is an AI autonomous company out there that I think is breaking the mold, it is Waabi. And to talk about what exactly Waabi is doing, which is very different than really anyone else, is the founder and CEO, Raquel Urtasun. Raquel, welcome to Freightways today. We're excited to have you.
Thank you. Really a pleasure to be here, Craig.
Well, let's talk a little bit about Waabi. It's such an interesting story. I wrote a book last year where I talked about Waabi as the idea, the simulations that you guys are running. Tell our audience a little bit about what makes Waabi special.
So what makes Waabi really special is on the technology side. We are bringing really a full AI-first approach, and the idea is that you unleash the full power of AI while still maintaining and enhancing the safety of the systems. Right? So this is: can you build an end-to-end system that is fully verifiable, that has really for the first time the ability to truly scale and meet customer demands, together with a simulator that is the same as the real world, so that you can really expose the system to any situation that might happen on the road and you don't need to discover issues on the road. Instead, before you touch the road, the system is so performant that it is superhuman.
So Raquel, one of my favorite parts of the Waabi story is what you can do in a simulation. I'm a pilot and I like to fly the simulators because it gives me the chance to sort of — and look, this is the reason that the aviation market is so safe and commercial aviation so safe, is because pilots are constantly training in these simulators. But one of the nice things about simulation is that you can do things you would never do in real life to really enhance your skill sets. And as I understand Waabi, really the thing, the aha moment for me, was that you're doing this for the AI itself, is that it doesn't have situational awareness that it's effectively in a simulation. It thinks it's actually on the roads learning these things. But what's great about that is you can create all sorts of situations that you would never create in real life that really can make the AI and the autonomous vehicle that much safer because it thinks it's in these different instances. Give us a couple examples of that.
Yeah. And it's very interesting that you actually have two AI systems playing against each other, right? You have one AI system that is the simulator, where basically it is creating the world around where the self-driving vehicle drives. And for the first time in the industry, it does it in a way that the self-driving system doesn't know that it's not reality, right? And we can actually prove mathematically that that's the case. So that's a massive enabler, as you can imagine, of both training the system as well as being able to prove that the system is safe in the first place, particularly on the safety-critical situations that there is no way that you can actually do this in the real world because consequences will be catastrophic, right?
And so there's a lot of new technology that went into building such a simulator. But think about from the child that jumps on the road, to emergency vehicles that go on the opposite side of the road, to a fire, an accident with parts all over the road, etc. All these things can be simulated in Waabi World. And since it's generative AI technology, it can create an unbounded number of situations itself — you don't need a human to be prompting the actual simulator. So it gives you all the coverage of anything that you will be able to see in the real world, while being grounded in reality and in physics, which is really amazing what this technology can do. And it enables the driving system to be super performant from day one, which is also very important as you think about development — you want to minimize the risk not only in deployment, the driverless, but also as you get there.
So I want to talk more about the performance of that Waabi driver, the AI driver. And recently in the news, you guys released an article, well, information about the Waabi driver previously trained on a Peterbilt seamlessly going into a Volvo, even with different sensors. So I want to talk about generalization. Can you explain what that means in AI and then how it's relevant here?
Yeah. And I think for the audience it will be worth to touch upon. There is several types of generalization that humans do seamlessly, but traditionally is difficult for machines. In the industry we talk a lot about generalization across environments and scenarios. And this is, for example, if you go to a new geography — this is important for our customers because it's the ability to open new lanes really quickly, right? That generalization is very important. Also across operational design domains — for example, you want as a driving vehicle the truck to not only drive on highways but also on surface streets, to be able to pioneer a door-to-door business model where since Q1 of 2025, we don't need terminals, we can directly go to the end customer. And that's a massive advantage and something that the customers really get behind because finally we have a solution which is tackling the needs of how freight is done today, and this is what they actually want as the self-driving product.
But the one type of generalization that is not really talked about in the industry — because nobody has this ability before this big announcement — is generalization across what we call embodiment. It's this idea that the virtual driver can actually transfer to new vehicle platforms. And this is a new type of vehicle that has different sensors, different ways of controlling that vehicle, and also different characteristics, different shapes, right? And the reason this matters for customers is that on one side, we want to provide a solution that is not just for Class A trucks. You think about box trucks, Class 6, 7, you think about robotaxis — where Waabi is also expanding, which is very exciting, a new vertical. But also when you think about building a global product — in Europe, trucks are very different, right? In Asia as well. So we have the ability to actually directly do so.
Also, customers typically have or prefer particular OEMs. So you think about how easy it will be with digitalization to serve multiple OEMs at once. And reduction of the BOM is also top of mind for everybody, right? So you want the cheapest, safest solution, and if somebody designs a new sensor that is cheaper or much better, it's super easy to integrate in the factory line. So this generalization across these different domains is tremendously important.
And the announcement that we made is that for the first time we were able to show full generalization across OEM platforms. And what this demonstration was, is that the Waabi driver was trained on Peterbilts, had never seen the new Volvo VNL autonomous, and without requiring a single training example, without requiring engineering, was able to drive perfectly both on surface streets as well as on highways, with parity with the Peterbilt truck. Imagine that — first time it goes into a new truck, you have parity with the rest of your fleet that is very different. So you can imagine what that means in terms of the scalability of our approach and the ability to serve markets, the ability to really serve what the customers really want.
Raquel, let's talk a little bit about the scenarios that you guys bring to your AI. So if not aware — and look, again, I'm a pilot in an airplane. I fly small airplanes and one of my favorite things to do is to go into the simulators, Microsoft Flight Simulator, and just try my hand at some of the most difficult approaches, or really the games that they give you. You guys are effectively doing that with autonomous vehicles. What are some of the scenarios that you really try to torment the poor autonomous AI inside of your thing? What are some of the scenarios that this thing would see that you would never put it through in real life?
Yeah. And to your point, one of the things that is very exciting about the simulator is that it has what we call adversarial mode, where basically it creates situations that are very difficult for that particular software release that you're trying to put on the road. So things that you can see — as we were developing features, for example, a couple of years ago when we were developing construction. Construction can be all different forms and shapes, and many things can happen, many different types. So it was very easy for the simulator to create many scenarios and then basically teach the skills of driving in these complex situations to Waabi driver. So this was instrumental in order to really get this feature to be performant in really just a couple of weeks versus what is years in the industry otherwise.
So that's one example where this simulation system was tremendously impactful. Where it has been very impactful as well now is as we prepare to the driverless launch with Volvo, is really on making the safety case and the safety assessment that our system is safer than humans. And in that it has had a tremendous role, and has allowed us to really understand the last bit of performance improvement, and has allowed us to actually get where we are today, which is on the verge of that launch with Volvo.
