Transcript: Developing a Versatile Autonomous Driving Business

Executive Summary

In this episode of The Road to Autonomy, Andreas Wendel, the founding CTO of Kodiak, unveils the technical strategy that has propelled the company to the forefront of autonomous trucking and defense applications. He details how Kodiak’s single, versatile software stack powered by a modular neural architecture can operate everywhere from highways to muddy fields without modification.

Andreas also explains the critical advantage of their modular hardware, including swappable Sensor Pods that reduce downtime to just 10 minutes, and discusses the rigorous, safety-first approach that enabled their landmark driverless commercial launch with Atlas Energy.

Key Topics & Timestamps

[00:00] Developing a Versatile and Safe Autonomous System from Day One

Kodiak’s initial approach was to build a “super versatile system” with safety as the number one pillar, based on the principle that “it has to be safe, or it’s not worth doing”. To achieve this, they hired an experienced team that understood which approaches work and which do not, and they created a framework to allow for safe experimentation to find the best solutions.

[02:25] Inside Kodiak’s single, Flexible Software Stack that Powers all Applications

Kodiak’s software is unique because a single stack can operate in all environments, from highways to muddy fields. This versatility is achieved through a “modular neural architecture” that allows engineers to add new “neural pathways” in parallel for different operational domains, such as defense or industrial use cases, without altering the core system. This allows for safe experimentation and the integration of the latest AI research.

[05:03] Balancing Data-Driven Learning with Verifiable AI and the “Rules of the Road”

Kodiak’s system is heavily data-driven, utilizing insights from over 3 million miles of real-world driving. This data is supplemented with established guidelines, such as the rules of the road for highways or known military doctrines for defense scenarios. Andy clarifies that while the system uses learning in every component, it is not a handcrafted rules-based system; the goal is to create a “verifiable AI system” that can be safely launched without a driver.

[08:40] How Different Environments (on-road, defense, industrial) Cross-Pollinate Learnings

Kodiak’s different business units are described as “additive, and not distracting from each other”. Over-the-road trucking provides millions of miles of data for hardware reliability testing, while defense applications offer structured testing scenarios for things like potholes, shock, and vibration. This cross-pollination of learnings was critical for the successful driverless deployment in the Permian Basin, which combines elements of on-road and off-road driving.

[10:45] The Competitive Advantage of Not Relying on HD maps

Since its founding, Kodiak has intentionally avoided using high-definition (HD) maps because they have poor “temporal resolution” and are difficult to keep updated. Instead, the Kodiak Driver uses basic map data for general routing and relies on its onboard perception sensors to understand the road as it exists in the present moment. This allows the system to react to real-time changes like construction or debris, an ability Andy states was essential for their driverless Atlas Energy deployment.

[16:10] The Modular Hardware Design behind the Swappable Sensor Pods

The Sensor Pod concept was driven by the customer need for maintainability and uptime. All sensors are concentrated into two modular pods mounted on the mirrors, which can be swapped out in just 10 minutes without requiring complex recalibration in the field. This design is a significant advantage for both commercial and military operations and the pods can even be interchanged between different vehicle makes and models.

[22:40] How Kodiak’s AI and Data-Driven Approach has Evolved Over Time

While Kodiak’s core principles have remained consistent, the technology has advanced significantly. The company started by avoiding HD maps based on the team’s prior experiences and has since evolved to navigate in areas with no maps at all, using overhead imagery to define a travel corridor. The company’s flexible AI framework has also allowed it to integrate modern technologies like self-supervised learning, auto-labeling, and generative AI as they have become available.

[29:15] The Rigorous Safety and Testing Behind the Driverless Launch with Atlas Energy

The driverless launch was preceded by a rigorous testing process grounded in a comprehensive “safety case”. This process includes extensive verification on the track, on public roads, and in simulation, where Kodiak can test for events that have never happened in the real world. The company uses a “live safety case,” continuously monitoring data to find and address potential issues before they become problems. This includes robust hardware testing for real-world conditions, like dust clogging a computer’s air intake, that simulation may not capture.

[41:00] Kodiak’s Strategy for Building a Profitable, Scalable Business Across Multiple Industries

Kodiak is focused on building a real business and is on a path to profitability, particularly through its industrial applications. The company also sees significant growth opportunities in the defense sector and in the vast on-road trucking market, where it can help address the driver shortage. The company’s overarching goal is to automate jobs that are “dull, dirty, or dangerous” and deliver a useful product that helps people.

Watch the Full Episode of The Road to Autonomy

Full Episode Transcript

Grayson Brulte: Andy, you’re the founding CTO of Kodiak. How did you first approach developing an autonomous driving stack? 

Andreas Wendel: Yeah, thanks, Grayson. Thanks for having me here today. And I’m really happy to talk about Kodiak today as an avid listener of The Road to Autonomy . it’s awesome to be on the show. So we started off looking into what can we do to build a super versatile system that of course, safety first, was, Something that we looked into from the beginning, it has to be safe, or it’s not worth doing. And so that was our first pillar of how we approached it. But then we also realized very quickly that you there’s lots of learnings out there on how to do robotics how to do autonomous driving So we hired a great team around this, not just experts in the field of all the components of the autonomous vehicle, but who’ve actually had this learnings and had actually seen, well, these approaches work well, but these other ones, even more importantly, don’t work very well, and then you need to go and build a framework for people to experiment in a safe way that they can actually go. And look into what works best. How should we approach this problem? So that’s how we have built up the company from the beginning. How I have built up the engineering team here. And, yeah, well, I think we’ll talk about how that led to a very fruitful solution.

Grayson Brulte: Versatility is fascinating because look, it’s paying massive dividends today. And they’ll continue to pay dividends going into the future. Why did you build versatility into the engineering program on day one? 

Andreas Wendel: Yeah. I would say on day one, we were really focused on over the road. So meaning on highway tracking. And that autonomous trucking quickly showed us, well, there’s, there’s lots of things out there that you need to handle from regular, ACC, so basically following other vehicles, to merge as that’s what people usually think about when you think of highway driving, but then there’s lots of things like potholes or debris. Or there’s lots of pedestrians on a highway as well. And we quickly learned that these are all things we have to handle. And that we need a lot of structure testing, a lot of other scenes and, to put in to our AI to get to a very robust and safe product. And that’s how we started to look into the versatility of that, collecting lots of various different data and making sure that we build a very comprehensive driver. So that’s how we got into this. And by now we can actually drive in various different industries from over the road trucking to the defense space to an industrial space.

Grayson Brulte: There’s, there’s flexibility, there’s flexibility in your hardware because you have different versions of the sensor pods, you have defense pods, you have your traditional sensor pods, but then I’m assuming there has to be software flexibility. Is there different stacks that for doing different applications, or how did you build flexibility into the Kodiak software stack? 

Andreas Wendel: Yeah. I think that’s actually very special about the Kodiak software stack. So. Our software can drive from highways, high speed on highways, all the way to roads, high speed roads with oncoming traffic, surface streets, dirt roads, all the way through a muddy field, right? And it’s all running on the same software. So how do we do that? It really comes down to the AI behind it. And, ,as I talked about in in the very beginning, it’s the room for experimentation that you need to create and safe experimentation. So we have what we call, a modular neural architecture, and that allows us to have several neural pathways that are in parallel. And if you, to, to make this more, More, more clear, it’s really neural networks that you can add to this architecture that actually go in parallel, which allows you to always use the latest and greatest of the state of the art of research. We go, we test it. We make sure it’s safe. We make sure it can be verified. On the road, but especially in simulation, we run lots of simulations every day. We take these vehicles to the track. We see how they perform in various different ODDs. And as we expand our ODD, the operational design domain, and go from on road trucking to defense to industrial use cases, we can just add additional neural pathways to it and make sure the system is safe. And so that framework of experimentation, that ability to add to the system is incredibly powerful and helps us achieve that versatility with one and the same software.

Grayson Brulte: I would put a neural pathway in simple terms. Are you just downloading Encyclopedia Britannica? So if I sat there and read the whole Encyclopedia Britannica and became an expert on say military history or a different type of history for you to become an expert in that driving, is that what you’re doing? Just giving the system more information to become an expert in that domain, whether it’s off road or whether it’s defense.

Andreas Wendel: I would say there’s lots of data in it. So it’s, we drove more than 3 million miles at this point. So we have seen a lot of things out in the world and we absolutely use this data for them. Every part of our system. So it’s a very data driven system. However, there’s also rules of the road, right? There is best practices in how you drive on highways, but also in defense, there is actually, a lot of known doctrine on how you actually approach certain situations. And so you want to have a method to get that information in as well. That’s where maybe where the encyclopedia comes in. It’s both data that you see out in the world, but also guidelines around it. Now that is not to be confused when people say like, ah, this is kind of a rules based system. It’s all handcrafted. That’s not at all what we do, right? We use learning in every single component of our system, but it’s important to, , to have a verifiable AI system to actually be able to launch it. And, as I’m sure you have seen, that’s, that’s what we’re doing. We’re actually launching these systems driverless. On the roads today.

Grayson Brulte: Is that approach what’s allowed Kodiak to scale safely across multiple domains? 

Andreas Wendel: Yes, absolutely. We do build up a safety case that consists of several different pillars. So 1st, you need to have a safety culture. That’s very important. If you don’t have the safety culture, it’s hard to make decisions and come to conclusions because you can’t monitor every single person every day. And so really having that culture is crucial. Then you want to design your system to be safe. And there, there’s a lot that goes into this from functional safety over kind of what is, what is known as SOTIF, which really says, well, how is the behavior that you do safe? And we do a lot of, of risk analysis there. There’s a lot of probabilities involved there in figuring out what is actually a safe thing to do. And then you go and say, well, now I’ve designed it, but I need to manufacture it in a safe way. I don’t need to operate it in a safe way. And I, go and make sure that it’s from a cybersecurity point of view, that it’s safe. And finally, you want to monitor your system, because you will always, as you drive more and more, you will encounter new situations that you might not have seen. You want to make sure that you stay safe, even in these situations, and even as your environment changes.

Grayson Brulte: I was listening to a really interesting podcast last week with Marc Andreessen and Ben Horowitz around AI and Mr. Andreessen talked about hallucinations around Gen AI. And if you look at Gen AI, Especially one of your competitors in the field is everything can go end to end without any let’s call it, let’s use the Eric Schmidt term adult supervision, because going back to your Alphabet days, it seems that you need to have that oversight in order to successfully do this, because as Mr. Andreessen said, the system can hallucinate is that we’re doing, you’re allowing the neural networks to work, but you’re providing that supervision to make sure they stay within their, their guidelines within the safety framework? 

Andreas Wendel: Yes that is absolutely correct, We do use generative AI in various different components, but being able to build a verifiable AI, right, that you can then build the safety case around is incredibly important. And our Gen AI is really around LLM, LLMs, but we also use VLM. So basically fusing in visual data as well to make sure that we can recognize scenes that we can describe, but maybe have never seen. And since we have so much data out there, our own data, but, but also other data, that we have access to, that actually gives us a huge, , foundation for all of our models and, lets us build a really robust and scalable AI.

Grayson Brulte: How do you test the AI for, if you look at the recently announced Atlas Energy Deal, congratulations on that. That’s running in the Permian Basin. There is heat, there is wind, there is dust. And if you look at your defense with Department of Defense, with the DIU Defense Innovation Unit, there’s mud. There’s all sorts of other elements there. And then if you’re on road, as you’ve been very kind to put me in your truck, we saw people walking in the highway in Texas. I saw somebody walking a dog, somebody riding a bicycle, and we’ve seen debris in the road, your system was able to handle that. How do you train for all these different scenarios? Since the environments are, those three environments are completely different.

Andreas Wendel: Yeah. It’s, it’s one of the beauties that these parts are really additive, and not distracting from each other. So. We see that they basically cross pollinate, right? Like we actually go and we run over the road and that gives us lots of miles. It lets us test the hardware, see that, we actually have the reliability over many hours of driving, many miles of driving, and then we take the same hardware and we can talk about that as well. It’s actually very modular and really testing the same things in the different use cases. But we take that same hardware and we run it in a defense setting. And, we see lots of potholes. We see lots of shock and wipe. We see lots of heat, in various different environments where there’s bushes and lots of obstacles that you usually don’t see in a highway environment. And so putting those two together is actually what usually. People pay a lot of money for to see one or the other people would go and say, Hey, I need this structured testing and I have an interest structured testing campaign. Well, we can do it as in the defense environment and actually, use those learnings. And then people also go and say, I need to do accelerated life testing and get many, many miles on something. Well, we get that from the on road. trucking part And so when you put the two together, actually this Atlas deployment, in the Permian basin where we now haul, lots of frac sand and where we have deployed driverless, we just announced that recently, that is actually, a great application for those two get fused together because what you need to do when you go to various different well sites is, they actually change. You need to go to various different locations. It’s very sandy and dusty out there. So you might not actually exactly see where the road is every day. , and it’s, it’s something that really our stack has been built for from the beginning.

Grayson Brulte: You look at, also look at the Permian basin, you’re running on dirt roads. You’re not running on paved roads with markers that you’re not, are you relying on mapping technology? How, how are you doing that? 

Andreas Wendel: So from, from day one at Kodiak, we did not use HD maps. And maybe to explain to listeners who are not familiar with that, many companies go out and they create a very detailed picture of the world at some point, and that you could say that’s a very high spatial resolution, but very low temporal resolution because you do it once. And then you usually don’t update it for a long time because it’s hard to update it. It’s a lot of work to create these maps and Kodiak takes a different approach. We actually use a very, Maybe low resolution in terms of spatial. And we just say, Hey, how do I get from A to B kind of more like, like you would imagine your Google maps, and technology work and, and how to tell you what, what the route is, we know for on a highway, where the exits are. We know. in these, other networks, like where you, how you want to go from point A to B in a defense setting. And then we go and drive that, but we use our onboard perception to say, where should we drive? How does the road actually look today? Did the highway change? Did someone come and move? The lane markers, where is the drivable surface in such a sand environment? Is there, are there lots of piles of sand on the side of the road that we shouldn’t drive into? Right. Uh, it, it doesn’t matter that two weeks ago this was drivable. If today there’s a big sand bank, you probably shouldn’t drive through it.

Grayson Brulte: If you had to rely on HD maps, would you be able to do the deployment with Atlas Energy? 

Andreas Wendel: We would not be able to do that.

Grayson Brulte: Okay, so they’re chalked it up to a competitive advantage.

Andreas Wendel: Absolutely. Yes. I think both the not. Relying on maps, still using maps if, if they’re applicable, right? If we drive several times, we use maps as a prior. We pull them in and say, well, yesterday it looked this way. That’s of course, very useful. You don’t want to reinvent the wheel every day. But you use those maps as a sensor and a sensor can have missing data. It can have incorrect data. It has uncertainties. And so we take all of that, information in. And we look at, well, how much does it match what we see? And if it doesn’t match, and we’re not quite certain about where exactly to drive, we have to slow down as a human would.

Grayson Brulte: It’s a cautious approach. Let’s throw caution out the window here. Look at, if you’re operating with the military and theater, caution goes out the window, you have a mission. How do you deploy the technology in theater?

Andreas Wendel: Yeah. Well, we actually learned over time is that lots of military missions are not combat missions. You actually want to move vehicles. You want to reposition them. You want to get into an overwatch position. Where you might not want to be detected. And it’s not actually a direct conflict in a lot of cases. And so when you do that, oftentimes you actually move on roads. Is it dirt roads? It can be, two track roads. It can be actually through a field where you know it’s drivable. You also don’t want your vehicle to get stuck. Now the properties of where you can drive that matters a lot of what you’re transporting and what your vehicle is. So if you have a big semi, which the military of course uses as well to do logistics in the field and in theater, then you might want to go on different roads and different paths than if you drive. A pickup truck, like we’re using as well. We have an F 150 in our, , in our car pool, essentially. And, it’s very different also than if you use a tank, right? You just can run over different things. You take different routes. And so our system has actually been designed to handle all of those and take the . the properties of the vehicle into account. And we learned that it’s really important to give, the Department of Defense a way to plan those missions, to say, what, what are the properties of the vehicle? What do you want to achieve? Does it have to be fast? Does it have to be slow? And even, do you want to remotely assist or maybe even remotely drive a certain vehicle? And that has been extremely useful in all the interactions we have, we have and had with the Army.

Grayson Brulte: The Army is a fascinating logistical enterprise. The men and women, in harm’s ways, I thank them for their service, but if you look at it from a logistical standpoint, The Army moves plenty of goods over our nation’s highways from from base to base. As Kodiak expands the relationship with the Department of Defense, perhaps could you go from a base in say southern Texas to Oklahoma and start running logistical routes for the Army using your over the road technology as in addition to your off road technology? 

Andreas Wendel: Yes, absolutely. And that can scale to a worldwide environment as well. There’s lots of semis being used in the army, lots of goods to be transported. And I think, some people even say wars are, are one, not, not directly with weapons, but they’re actually one with logistics. And I think that’s where we can actually help And, make sure that we stay at the forefront there.

Grayson Brulte: You have to make sure that the front line has all the supplies that they need because if you run out of water, you run out of ammo. You’re in deep doo doo. There’s no other way to, to say that. See, that’s the software side of the Kodiak stack. Then you’ve got, I’m a very big fan. I’m an unbiased fan because I think it’s absolutely brilliant. You have Sensor Pods. And you have the Defense Pods for defense. And you have the traditional sensor pods, which are cool. And I’ve done a pull up on them. I couldn’t believe that actually held me. and why was the sensor pod concept developed? 

Andreas Wendel: Yeah, I’ll maybe just explain to, to our listeners what a sensor pod is and, and what we did there. And so the very beginning, when we started Kodiak, we said, well, how do we make sure this is maintainable? How do we make sure that . a truck that goes across several different states and needs maintenance every now and then, or even has a sensor fail at some point, because of course that will happen and no technology is perfect. Then how do we actually get that replaced? Because the number one . thing we learned from our customers, talking to them every day was like, well, these trucks need to be on the road to make money for me. And it goes the same for us. Like if our trucks are not on the road, then they don’t make money. Now, it turns out this is the exact same thing in the defense space, right? If you’re out in theater, you want to be able to replace your sensors quickly. You don’t want to go out and have to calibrate it or needing a turntable or something like this. You need to be able. To get those vehicles up and running and keep them up and running. And so because of that, we said, well, how do we position our sensors and all of our hardware? And, we found it to be extremely modular to put all of them on the mirrors. And so that’s really what the sensor pods are. They are basically two mirrors. If you look at our, our semis, you would probably say like, well, they’re pretty big mirrors. And the reason is because they have to include the human facing glass, still, that will go away over time. So they will shrink considerably. And the defense pods already don’t have that anymore. So the defense pod are noticeably smaller. So. They had, there are slightly different form factor to integrate on various different vehicles. But in general, the concept is the exact same. You have all the sensors concentrated in, essentially two positions, one, left mirror and a right mirror. And that makes it extremely modular swappable. We can calibrate it in the factory and then just put it on. And, the swap of such a Sensor Pod only takes 10 minutes.

Grayson Brulte: If you think about this from what’s two different scenarios, one traditional trucking, let’s say you’re going Dallas, Houston or Fort Worth, El Paso on either end of that spectrum. Could you have sensor pods and it’s in stores or something happens? They could be replaced there.

Andreas Wendel: Yes, absolutely. Yeah. You can even give these sensor parts to someone who drives to a truck should something happen and actually gets it there. You can even put it into the vehicle, right? Now, you look at like how many times this has happened. It’s probably not necessary to have it in every single vehicle, but you could if you wanted to, and then you have replacement parts there when needed.

Grayson Brulte: And then in theater, you can have various, it’s called strategic locations or forward bases, different types of bases. You can have the defense pods there as well, and they can be changed out in theaters with the same amount of time, 10 minutes? 

Andreas Wendel: Absolutely. Yes. We have now integrated at this point into, I think, five different brands of vehicles across semis and like passenger vehicles and various other military vehicles. And, we can actually swap , the pods from one vehicle to the next. So it’s not necessarily dependent on which OEM this vehicle comes from. And that again is a huge advantage to, how many parts you have to have in, in storage, right? 

Grayson Brulte: I think about this, let’s, the maritime terms of you’re on a vessel and you have to coast guard rules, you have to wear a life preserver and they’re on the vessels, depending on the amount of individuals that are there on the vessel from, this is more of a design question, but just thinking out loud from a logistical standpoint, could your design team, so obviously you have the, the Kodiak compute stack, but could your design team integrate Sensor Pod storage there so no matter what happened, that vehicle always has a backup sensor pod on it, very similar to when you’re on a vessel, you have a life preserver. 

Andreas Wendel: Yeah, yeah, I think that makes a ton of sense. And we, we actually took this concept further, right? It’s not just the sensor parts. We essentially our, all of our systems, they have three components. They have the sensing part, which is in the sensor parts, and really modularized there. And then we have to compute, which is again, very modularized. So it happens that our semis in our F 150s use the exact same compute stack. So, we can swap between those, we can swap them out between those vehicles. And, that again, makes it very modular now that, and the third one, just to finish that thought, the third one is the actuation. The actuation depends a little on the vehicle, but again, it’s very modular and, it it’s make making sure that we have all the redundancy for that particular vehicle type and for the actuation that the vehicle has. So with these three components, we can. We can go and up fit various different vehicles and, make them safe driverless vehicles. Now, eventually those will come from a factory line, but even there, if you have less components, less positions where you need to integrate, that is actually a huge advantage to going to manufacturing and all of these modules can, Evolve independently over time because they have very clean interfaces to each other. So we see that as a major advantage, both for manufacturability, but also for keeping that maintainability in the field.

Grayson Brulte: The same compute sensor processor swappable to me. I’m going to go to Denmark here. You have Lego blocks. I know you’re from Austria, but we’re going to go to Denmark because you have you have Lego blocks here. How did you design that system from a technical standpoint to be? It’s called the Kodiak Lego block system. Because to me, if I’m purchasing your product or I’m using your product, Or I’m evaluating your product and say, Oh, I operate a mixed fleet and all this is swappable. To me, that’s a game changer compared to your competitors.

Andreas Wendel: Yeah, absolutely. That’s how we see it as well, right? It needs. Less, less training for many, many technicians out in the field. The training is much more simplified and, it’s way easier to actually swap it. And from a manufacturing point of view, it’s great if you can actually hone in onto these systems and not have to develop a completely different one. So again, that’s, that’s where I said before, it’s additive, not actually distracting in all of these different use cases, because suddenly we have different vehicles. We do the same, for instance, shock and vibe tests on our compute, on road and on defense. So suddenly our defense environment becomes a great shock and vibe test for the on road cases, right? So that’s, that’s how those really go together.

Grayson Brulte: The added defense, let’s go back six years to the founding of Kodiak. How is your approach to autonomous driving to your developing the stack changed? 

Andreas Wendel: I would say from the beginning, we looked into, well, what, what should we, what do we want to aim for? And we talked about not using HD maps now that in the beginning, for instance, we did this from the very beginning, but in the beginning it wasn’t bad. No one had ever done it, but we just had learned from prior experiences, from our team coming in from various different AV companies, that. It’s a pain to have an HD map. It needs many, many people to actually maintain it. And like I already said, the temporal resolution is terrible, right? You have to update it all the time and that requires a lot of work. So we said, well, how about we go and try to put as little into the map and to rely as little on it and to treat it more like a sensor. And so that is one of the concepts that we did from the beginning. But then over time, of course, we refined that. We can now. , use in the, in the defense space, for instance, we can run in areas where we have zero map, no map at all. We’ve never been there. And that’s of course a requirement for defense because you don’t usually know where you’re going, , in these cases to a detail of like, an inch, a square inch right on the ground. So what we can do there, for instance, is we can use overhead images and say, well, here’s your corridor. And that corridor can be pretty wide. And this is where I want you to go. Here’s your bounds. Don’t leave those bounds. And we send a vehicle there and it searches for the best path a lot within its bounds and arrives at its destination. So that, that is something we didn’t have on day one, of course, and that we have developed. And maybe similarly on the AI side, we have always used AI from the very beginning and we’ve always been very data driven, but we’ve gotten much more data for using , lots of self supervised technologies, and methodologies we use, auto labeling, these are all things and generative AI, as we discussed, these are all technologies that have popped up in the last years and have not been really the state of the art when we started Kodiak, but it’s important to give yourself this framework where you can actually put these technologies to good use and put them in so that you’re not static in your technology. And so I would say. It has developed a lot, but the fundamentals are all the same and the principles are very much the same.

Grayson Brulte: It’s evolving. That’s I think that’s a very fair statement because what you did was you built the technical foundation as technology evolved and advanced. What new technologies do you see on the horizon that could have a positive impact on the Kodiak driver that you could potentially integrate? 

Andreas Wendel: Yeah. I think especially on the sensing side, we, we do use, LiDAR. We do use radar. We do use cameras and we’ve never been particular about one of them being the main sensor. So all of the sensors really have equal weight in our system. We do a lot of early fusion where we take cameras and radar camera and LiDAR. We take all of these methodologies together and we make sure that if one of those fails, then we never have a failure of the complete system. That’s some of these neural pathways. They all try to not depend on everything so that if there is a failure in there, that you can actually still drive the vehicle. And so with that, I think that while maybe five years ago, lasers have been extremely critical. Nowadays, they’re actually still very important and useful, right? I don’t think today we could build a driverless system that is, that meets our safety standards and is deployable without the laser. But I think it’s going there, right? We can actually do lots of things. Ranging with only cameras and radar. , and so that’s maybe where the technology is going over the years that we can even further reduce the sensor set. I think we already have the leanest sensor set in AV tracking, but we can further reduce that. And that goes back, , to, to what the cost of these systems will be, overall and how you can further integrate it and mass produce it.

Grayson Brulte: I fully agree with you and there’s economic data. If you add up all the publicly traded LIDAR companies, the market cap combined is under $2.25 billion under that. I mean, that’s after the rout yesterday, but we’re still probably just over $2 billion. There’s not much there and it comes down. To cost. Do you see, is it going to be a depth sensing technology that breaks through a CMOS technology for camera? Will there be a, you see a new camera technology evolving that could potentially accelerate the move away from LIDAR? 

Andreas Wendel: I think it’s, it’s really in the AI that we see lots of improvements there where you can actually estimate a lot of this from prior knowledge, from, from understanding of the world of, of what the relative depth is to each other. So, that makes it even in the perceptions space, even more human like. And, we see that humans, of course, can only drive with, our eyes and that isn’t, we’re not yet there, right? That’s not yet something that, that I would sign my name under and say, yeah, I can go driverless and it’s actually as safe as a human. , but I think it’s going there, over the next couple of years, we will see a push towards this. And I don’t think we need to invent new sensing technology for this. But today we do need LiDARs. We do use radars. We do use cameras and they’re all incredibly important to meet that, that very high bar to actually launch, in a driverless setting and many people talk about how do I build 10 thousands of it? , and that’s very important, but you first need to launch your first truck. You need to, , launch your first hundred trucks. And there’s a lot of operational complexity, a lot of. , we call it the Kodiak driverless launch protocol, right? Like how do you actually roll this out? How do you get it t to the operations people on the ground to actually run it day in day out 24 seven And have a good experience meet your SLA, right? It’s really going back to to an actual business and that’s what we’re doing with Atlas We’re running an actual business we ran, this driverless, run that we already did was a commercial load where we delivered sand and that’s just incredibly exciting.

Grayson Brulte: You are business. I’ve said that to Don Burnette many times. Kodiak is truly business and it’s one of the main factors why Kodiak is number one on the leaderboard for autonomous trucking and you’re also number one on the leaderboard for defense because yourself, Don, the entire executive leadership team and the board have had the privilege of speaking with him many times. We’re a business. We’re a business. We’re a business. And that’s really positive to see. So you’re driverless now with Atlas in West Texas in the Permian Basin. What went into that launch from a technical standpoint? Was there the technical checklist that you alluded to that went into that? Was it rigorous safety testing? What went into allowing you to launch fully driverless with Atlas in West Texas? 

Andreas Wendel: Yeah, absolutely. We have a very rigorous test process for this and, , we definitely don’t take that lightly. Like I said, in the very beginning, safety is our number one priority there. And, , it’s not worth doing things that are not safe, right? , that’s what our customers expect from us, but that’s also what we ourselves expect from the technology. So what we do is I already talked about the different pillars of the safety case, but you go and do a lot of testing on that. You do a lot of verification, validation on the track. On the road, lots of miles that we drove, but especially in simulation and in simulation, you can go and look at things that you have never seen that have never occurred because even if you look at 3 million miles of data that we have collected so far, that’s not enough, right? There’s still lots of . things that can happen in the, in the real world that you have never done. So that’s where you go into doing really a top down analysis. It’s what, what is often referred to as a safety case. You go and look at how likely are certain things to happen. What’s the severity of them if you don’t perceive them or plan around them correctly. And then you go and check those off and you say like, well, is this actually a risk that is still there and that I need to mitigate more? And you go and put mitigations in place for those various different risks and you check against it. Now, this sounds like a one off process. You do this once and then you go launch, but that’s not how it is, right? You actually have to live monitor this. So we call it a life safety case where you take all the data that comes in and then refine it and you say, well, are there any surprises in my data? Is there anything that I didn’t model or that I didn’t fully understand? And even there, you don’t want to wait until say a collision happens. You want to see some of the early indicators of that. And you want to measure those, and make sure those actually get flagged. And our triage teams, our simulations teams pick up on this, and that creates priority work items for us. So that’s kind of the, the flywheel there, both in terms of machine learning. We have a, an active learning flywheel where we take that data and continuously refine it. Okay. That’s incredibly important. That is actually what makes these systems work, but not just on a machine learning side, right? Like you want, want to know this about the hardware, about how good is your hardware? Do I see any, anything getting, getting loose or not actually working together the right way? Do I see any, any bit faults on a CAN bus? And I think people often. Take this as like, well, I can do it all in simulation, but obviously hardware is incredibly important. And we have a system that needs to work reliably and redundantly. And it’s just, just, uh, one, one more example here. People often think of redundant breaking as being super important. And of course it is. So we have triple redundant breaks. But if you go and say, well, My compute could also overheat because dust was collecting in the air intake. Right. As an example. Well, that’s not as obvious in simulation. That’s not as obvious to come up with and say, like, well, you have to actually fix this. But it’s a very simple maintenance item in, in various different systems that we employ all across the world. So these are learnings that you have out in the field and that you, that are actually incredibly necessary to launch this technology safely.

Grayson Brulte: You see, this is fascinating because You hit the nail on the head. This is me perspective. I don’t think a simulation only approach works from a purely for a lot of variety of reasons. I go down a three hour conversation with you. Why? It’s a tool, but it’s not the only tool when you’re pressure testing it. You look at sands. You’ve run through all these crazy elements in a West Texas It can get hot. You can cook an egg on the road. It gets it gets so hot. Do you run these trucks manually through all these different scenarios and plan for what if or say, Wait a second. What if we blew dust in here? What would happen to plan from all this from its real world practicality? We’re not you’re not operating in a clean lap at the end of the day. There’s elements there. There’s birds, there’s winds, there’s all sorts of stuff. And it hails. So you, do you run through all these different real world scenarios outside of simulation to learn and understand how the truck would adapt? 

Andreas Wendel: Yeah, that’s absolutely right. Like temperature there, there’s great temperature chambers that you can put, your hardware into, right. Where you actually run through temperature cycles. So we have an incredible hardware team that goes and tests the autonomy hardware. We also use trucks that are built for this and trucks are really an incredible, incredible vehicle, right? Like, building these trucks over years and years and refining them. They’re incredibly hardened. But they still break every now and then, right? They need maintenance. And so learning about what you actually need and what you need to do when that makes what your heart makes your hardware really work extremely well . but before that you go and test it to that, you go test it with temperature cycles, you do accelerated life testing, and you make sure that you can withstand those different conditions. And that can be from, from rainy conditions, from very cold, snowy conditions, to super hot, , conditions. And, , that’s what we run our vehicles through.

Grayson Brulte: From the hardware perspective, especially from the compute and you’re running and we’ll keep going back to West Texas, or you’re running a military, do you have to build different types of cases? If you look at it in a military application, potentially a bad guy could shoot out the GPS could shoot out that. And then in West Texas, you have the heat issues that you’re looking at liquid cool, potentially running in West Texas while and the defense, perhaps there’s some modified defense version of the case. How are you building your enclosures for the different environments that you’re operating in? 

Andreas Wendel: Yeah. I would say at the moment, they’re very similar still. And they, you know, it’s, it’s more about the redundancy in there. If someone really shoots out your sensor pod, I don’t know that the enclosure really helps you that much. And so that’s where you actually want to say like, well, if any particular part of the system fails, I want to be able to detect it and react to it and reacting to it. For instance, in an on road and a highway case, it usually means to pull onto the shoulder. It means to pull over, make sure that you’re clear of the That the regular driving surface that you’re not blocking anyone else. Now, there’s cases where you might actually just have one lane and you actually stop there, then you want to stop in lane. But again, you want to then be able to give a call to your 24 7 operations center, which we have, and notify someone who can then, depending on what the issue was, come pick up the vehicle. But ideally, You don’t even wait for that because you will be blocking the highway, right? You actually have means to get this vehicle onto the side of the road through remote assistance. And that’s really powerful in these, environments as well. Now we talked about the sensor pod being essentially say one of them is actually damaged that in a regular on road case. That might actually be a problem that you don’t want to recover for from and you want someone to come service the vehicle, but in a defense use case, depending on what your mission is and what you’re doing, you might actually want to continue and that’s really down to the commander in the field to say, Hey, What is my mission? What do I actually need to do? And I think we had a lot of learning now working the last few years with the army about this to say, how can we support our customer? And that’s again, one of the pillars of, of Kodiak from the beginning is we’re going to optimize for what our customers need, be it on road where we work with, various different. Large companies, large carriers, but also shippers, across the U.S. And we work with them very, very closely to see what do they need. And the same goes, in the defense space. And the same goes, of course, with Atlas, who are, a very valued customer there in the Permian Basin.

Grayson Brulte: You’re optimizing for your customers needs, which you mentioned this environments. What environments can Kodiak operate in today? And what environments do you see Kodiak operating in the future? 

Andreas Wendel: The environments, we call it operational design domain. They’re very far stretched for us because we have such a versatile system. So from highways all the way into the field, as I said before, and there, of course, all of these need hardening, , in this case, and you go after what the actual use cases we started with. The technology in, in the Permian basin now where we have actually been able to take the driver out, show that we’re safe enough to be driverless. And so that has been an incredible push to get that out. And it’s actually very. It’s a really big stepping stone for building trust with people as well, because it’s hard to for many people who are not as close to the technology to see from day one that yes, you build a driverless safety case and you run it at 65 miles an hour on the highway with a big semi and 80,000 pound vehicle. And building this up over time, showing that, look, our redundant actuation actually works and it is safe. It is, , the use cases with oncoming vehicles in the Permian, and actually hauling sands, maybe at lower speeds first, but. Having still to react to all the other vehicles, to pedestrians, to vehicles that I had never seen before that exist in oil and gas fields, and they’re, they’re weird looking vehicles, but still, our, our truck needs to react to them properly and safely. All of this builds up trust. And I think that’s incredible. Incredibly important for our industry to build that trust, to show people that this works because we can run mathematical models and do analysis and simulations all we want. It’s really to gain the trust of the general public and the people who use it. That is the critical piece in the end. And, many people are not as mathematically inclined that they just trust, , a large probabilistic model.


Grayson Brulte: Without trust, there is no autonomy. There’s no warehouse autonomy. There’s no autonomy in the QSR quick service restaurants. There’s no And Class 8 Trust is that magic glue that holds it together. When it’s happening, it’s building. The public is being receptive. We’re seeing really good polling data. The public believes in this technology because Autonomy helps to fight inflation and autonomy is good for the economy and it’s going to usher in what we call the autonomy economy. Kodiak today is operating on Class-8 trucks. You’re also operating as a Ford F 150 through your DOD Department of Defense programs. Are there any limitations to the vehicles that the sensor pods could make fully autonomous today? 

Andreas Wendel: I don’t think so. We, we always, for every vehicle, of course, we look at what’s the field of view requirements, right? We have to go and look if, if the vehicle is weirdly shaped and articulated, of course, we optimize what the Sensor Pods see and, and, what is necessary to see so that we can show safety at the end of the day. And, that is the only constraint on it. But other than that, the sensors that we have there works incredibly flexible in terms of what exactly we use. It’s often doesn’t come down to the exact specs of a sensor. It comes down to, what the reliability of the sensor, the lifetime of a sensor is right. You can combine various different sensors to achieve your goals . but if you. Make sure that that is, that that is true. Then, we can put our Sensor Pods into various different use cases.

Grayson Brulte: Different use cases equals, which I said earlier, a diversified business. Kodiak from day one was built to be a business, not to be a science project. Andy, how do you see the Kodiak business growing over the coming decades? 

Andreas Wendel: I think we’re on an incredible growth path now, with industrial actually, , being an area where we can make the entire company profitable. That is just, something that is very new in the AV space. And that’s something we’re incredibly excited about and where we are now at the forefront of deploying this driverless technology, but of course the sky’s the limit, right? Like there’s lots more that we can deploy in the defense space. We already got a $50 million dollar contract to develop this technology, but that’s only the beginning, in defense. You, we can scale this much larger. And of course, again, we want to deploy this as a product. So we’re working very closely with the department of defense of making that a for making that a reality. And then, the biggest space of course, here is on road trucking where there’s lots and lots of trucks, , that we can actually deploy and help people by doing so, because we all want their goods to be delivered, and, have them at, at home. And there’s less and less people who actually want to do long haul tracking, short haul, logistics is something where the jobs are local, but if you do long haul, you’re on the road for very, very long, , times, very long durations, and many people don’t want to do this anymore. So to summarize this, I think, Kodiak is ready to automate, any jobs that are, are dull, dirty, or dangerous. And that’s really what we’re focusing on. That’s where we want to help people with. And I think there is an incredible amount of . of industries and technologies out our industries and areas out there where our technology, can help a lot and we’re committed to making an actual product that helps people rather than just throwing a technology over the fence.

Grayson Brulte: And I want to add this. I said this before, you’re building a business, you’re helping people, but you’re also building a business that’s going to reward your, your shareholders. Because if you look at the global economy for very different applications that Kodiak’s involved with or not involved with, we have growing labor shortages. There’s growing labor shortages in the fast food industry. There’s growing labor shortages in over the road long haul trucking. There’s growing labor shortages in forklifts. And I read an interesting thing today in the Financial Times. There’s a growing shortage of masseuses, people that give people massages. And so now that’s getting automated. There’s in New York now, there’s an automatic autonomous, fully autonomous robot that gives you a massage. So there is growing labor concerns and automation is going to fill that void. If you look at the Kodiak expansion into industrial, there’s a lot of job shortages there. You’re going to fill that void in your customers and be able to build big, profitable businesses because technology that you’re developing. When you put Kodiak in defense, you’re going to allow the United States Army and Military to conduct the missions that they have to conduct because you’re keeping our men and women out of harm by creating really great Technology and over the road you show up the supply chain. We all saw what happened in COVID, my opinion, if we had autonomous trucks, the Port of LA San Pedro, we wouldn’t have had the backup we had. And that’s what we need. Your technology and your development at Kodiak is good for the economy, the U S economy and the global economy. Andy, it’s always wonderful to have you here. And this has been fascinating as we look to wrap up for today. What would you like our listeners and viewers to take away with them? 

Andreas Wendel: Absolutely. Thanks so much, Grayson. And this was a great conversation. And for me, , you, you said it perfectly. It’s. We’re building a real business for, , really moving this toward profitability, which has been something that has been missing in, in autonomous vehicle development and technology for a long time. So it’s really about building a real business that adds. That adds safety, adds, productivity, and make sure that we can help people, with it in all of these, , different areas. And I think we’re, we’re on a great path to this. Kodiak is growing, so we’re always happy to talk to, to new customers, to new team members, and get everyone. In who can help us and who wants to help us on this journey. It’s an extremely incredible and exciting time to now really have launched this year, the first driverless semi truck in a commercial application and, yeah. Stay tuned. We’re going to launch more driverless vehicles, different types this year. And, yeah, it’s going to be awesome.

Grayson Brulte: Kodiak’s going autonomous and Kodiak’s building a business. The future is bright. The future is autonomous. The future is Kodiak. Andy, thank you so much for coming on The Road to Autonomy today.

Andreas Wendel: Thanks. It was a pleasure. Anytime.

Key The Road to Autonomy Episode Questions Answered