Daniel Abreu Marques: Welcome back to Autonomy Insiders. I'm your host Daniel, and my guest today one of the more contrarian companies in autonomous driving. Igal Raichelgauz is and CEO of Autobrains, an company that argues the whole industry has bet on the wrong kind of AI. So Autobrains landed two big announcements at Nvidia's GDC in Taipei,

igal: Welcome back to Autonomy Insiders. I'm your host Daniel, and my guest today one of the more contrarian companies in autonomous driving. Igal Raichelgauz is founder and CEO of Autobrains, company that argues the whole industry has bet on the wrong kind of AI. So Autobrains landed two big announcements at Nvidia's GDC in Taipei, ⁓

Daniel Abreu Marques: So the timing couldn't be better. Igal, happy to have you on the show today.

igal: ⁓ So the timing couldn't be better. Igal, to have you on the show

Daniel Abreu Marques: So to start, maybe set the scene for us. What is autobrains and what do you actually do and what problem are you trying to solve?

igal: So problem we are trying to solve is to bring autonomy to every car, right? So we have today already quite a lot of autonomous cars. They're called Robotaxis We have thousands of them. And we also have millions of cars that have ADAS features, those L2 L2++ features, which basically allow car to drive to some in some areas, in some scenarios, with the supervision of a driver. So you must keep your eyes on the road. And so our goal is basically to bring the autonomy to about 1.5 billion cars on the road, regular cars, and to bridge these two gaps between autonomy and scalability.

Daniel Abreu Marques: And in one sentence, what does autobrains do differently from everyone else in the autonomous driving industry?

igal: And in one sentence, what does autobrains do differently from everyone else in the autonomous driving industry? I believe we are taking the technology one step closer to how the human driver and the human brain works. Specifically, we call this or this additional technological jump agentic AI, where we take the complex task or the complex problem of autonomous driving. And we let the AI to break this problem to large but finite and manageable number of smaller, narrower problems, where each of those problems is solved by a separate agent. Think about a specialized driver for certain conditions, ⁓ agent is being ⁓ by an orchestrator on the fly when the car drives. And so the benefit of this approach. is in two major directions. The first one is specialization. Specialization results and better accuracy, better safety, and better performance. And the second is ⁓ Instead of building a single monolithic huge end-to-end model, we break the problem to those specialized agents where each of them is activated in different areas and different times. ⁓ And each of them is dramatically smaller than the big one model, consumes less resources, less computational power, and results in less cost.

Daniel Abreu Marques: sense and you've branded the system also as I think liquid AI then skills and now ⁓ agentic AI with specialized driving agents ⁓ that switches on only when needed as you ⁓ explained. ⁓ is that the new architecture each time or was it more like new language for the same core idea?

igal: Makes sense and You've branded the system also as I think liquid AI, then skills, and now ⁓ agentic AI with specialized driving agents ⁓ that switches on only when needed, as you ⁓ explained. ⁓ is that the new architecture each time, or was it more like new language for the same core idea? It's the same approach. You know, the approach of skills, when we started the approach of skills and the evolution of this into agents, ⁓ basically is the same approach. So each agent has a certain number of skills. So when you have a driver, the driver still has to have the skills like an emergency situation or overtaking a car or passing a busy junction. All those are different skills of these agents. So this is basically an evolution of our technology. The liquid AI architecture essentially is the ability of the system to dynamically orchestrate those agents. And company really evolved from We started our journey six years ago when the industry was based on separate building blocks, right? So there was a perception building block, a path planning, ⁓ and an actuation. And although our technology was always end-to-end, it was always neural network-driven and data-driven and learn-based, learning-based rather than rule-based or compound-based. at the time the industry was in this ADAS-like architecture, so we focused on perception. ⁓ And now in the last ⁓ year, we really scaled this technology all the way to end to end stack, which allows us full autonomous driving.

Daniel Abreu Marques: And the the industry has mostly also converged on the the end to end model, or that's one model that's ⁓ often discussed. and you argue the opposite with the many specialized agents. So ⁓ why do you think the consensus is wrong?

igal: And the the industry has mostly also converged on the the end-to-end model, or that's one model that's ⁓ often discussed. and you argue the opposite with the many specialized agents. So ⁓ why do you think the consensus is wrong? You know, interestingly enough, if you look at the broader AI area, especially in the last year, year and a half, the AI shifted from a monolithic approach to agentic AI, right? Agentic AI today is almost a buzzword out of autonomous driving, where it's very clear that if you want to solve problems, you must have on top of a general ⁓ many specialized agents. And you know, the leaders ⁓ ⁓ like Jan LeCun and many others. of them recognize and basically advocate for specialized AI rather than general AI. There is no such thing as general AI. Even human brain is ⁓ kind evolved for a certain set of functions. And so agentic AI today the mainstream, ⁓ it's not mainstream in autonomous driving, and we have an opportunity to innovate here. The technical reasons end-to-end monolithic approach is stuck today. ⁓ in its ability to scale comes from you know the basic scientific foundation when you train a single model you have ⁓ essentially one to optimize the error the average error of this model and so when you have a lot of edge cases The only approach you can use in order to solve this edge cases to collect more data and to continue training. And these edge cases are usually a very small portion of the training set. So the process gets into this area of diminishing ⁓ Yeah, the term was used about one and a half years ago about LLMs before there was a shift to agentic AI And so that's what we see in the industry. You see it in many for for for many players. Sometimes there are recalls, sometimes there are retraining of the system, but there is a race to collect more data. It's all trained, there is no knowledge distillation, and the models grow. And so if you want really to get this match between scalability, affordability and autonomy, we believe the agentic AI approach is the only way to go.

Daniel Abreu Marques: And you mentioned edge cases and this is everyone's headline problem. ⁓ can you give us maybe one concrete scenario where your system handles an edge case maybe today that a conventional ⁓ stack cannot?

igal: Sure, you know, so many scenarios ⁓ that see today Waymo are have challenges, right? So, for example, adapting the the behavior of the car around the school bus. understanding the intent of pedestrians, differentiating between a police scene and just you know a bunch of cars, understanding if there are pedestrians, let's say an adult holding a a kid, ⁓ whether the adult Does does hold the kid or not? If there is a a certain, let's say, a ball on the street, whether a kid going to pick it up. So all these scenarios where you need more common sense beyond pixels, where you need really understanding of the situation and not just the pattern matching and reactivity to the situation, that's where our technology has a very significant advantage in the ability to think and reason and not just immediately react.

Daniel Abreu Marques: Super interesting. And you also claimed I think roughly ten times less compute and almost no labeling. So where does that efficiency actually come from? And how do you prove it also, maybe to a skeptical OEM safety team?

igal: Super interesting. And you also claimed I think roughly 10 times less compute and almost no labeling. So where does that efficiency actually come from? And how do you prove it also, maybe to a skeptical OEM safety already hit the roads with these cars on a very massive scale. And it runs on very low compute. We are talking, it depends on again on the product, but even the more extreme high-end products, are talking about tens of tera operations compared to thousands of tera operations, so two orders of magnitude. And the reason ⁓ for achieving is all about breaking the complexity. When the problem is complex, you run into this exponential scale of the model. Hundreds of millions, if not billions, of parameters. If you break the problem to simpler independent problems, and that's where we believe the intelligence is the ability to take a problem and to divide it into simpler problems that can be solved and then maybe divided again. That's the recipe by the complexity, reducing the resources. Super interesting and

Daniel Abreu Marques: Interesting. And one other thing that is also always a a weakness or a problem with these end-to-end AI systems is also traceability. So how does your approach make a st a single driving decision explainable to a regulator or an auditor?

igal: One other thing that is also always a a weakness or a problem with this ⁓ end-to-end AI systems is also traceability. So how does your approach make a st a single driving decision explainable to a regulator or an auditor? Yeah, so so one of the big challenges in monolithic end-to-end is the lack of modularity. And modular approach is very important, you know, there is no contradiction between being neural network or brain-inspired and being modular. In fact, if you think about it, end-to-end is a single neural network. While if you think about the human brain, it's not a single neural network. It's kind in the early days of AI, it was called a society of neural networks, where each neural network we call it an agent, but those neural networks and agents they communicate, collaborate, compete, and reach a certain consensus and decision. And this ⁓ ability to build modular system has a lot of practical benefits. Some of them relate to explainability, so you actually can localize to a certain agent a certain functionality and to a certain skill, a certain behavior, to a certain tool, a certain utility that's very similar to what happens in AI today. But beyond that, the whole cycle of training, validation, homologation, instead of being executed. On a system level, which can result in tens of millions of dollars of data collection and cloud and so on. This is being executed on a modular level, as modular as it can get to a tool level or a skill level, and not no need to basically redo whole process on the level of the entire system.

Daniel Abreu Marques: Super interesting and I think we we covered a lot of good stuff on the how how and maybe let's talk also about your your business model. ⁓ maybe ⁓ let's co get co get concrete. So what does autobrains actually sell and to whom? And so basically how do you make your money?

igal: That's super interesting. And I think we we covered a lot of good stuff on under how. And maybe let's talk also about your your business model. So maybe let's go get con get concrete. So what does autobrains actually sell and to whom? And so basically how do you make your money? So we have today three lines of business. The first one is r basically relates to mass production programs. ⁓ It starts from the most basic functions like emergency brakes, adaptive cruise control, lane keeping that ⁓ already production ready and being shipped to the road, to the roads globally. and we sell this in a very basic license fee. For us, this ⁓ direction is super important beyond revenue generation. We prove that our technology is automotive grade That's a very high bar. There are many companies today. That show the technology on a scale of ⁓ hundreds of cars, even when you get to Robotaxi the overall market is thousands of cars. When you get to mass production, we're talking about millions, and that's a very different ⁓ data set, they very different requirement. So, this is a very important ⁓ business ⁓ vector for us, the mass production programs. The second one is the Robotaxi it's an existing ecosystem. We have joined relatively late to this ecosystem, about a year ago. And as you know, we have announced both a partnership with Vin Group with Vinfast and Nvidia to scale Robotaxi in South Asia. We believe it's a very interesting geography for us, very chaotic driving, and very challenging. So, probably the highest bar in terms of performance. And so, on the other side, in Munich, in Europe. are working ⁓ ⁓ in collaboration with Uber and Nvidia to scale multiple with multiple OEMs, Robotaxi on European streets, where the regulation and safety is the highest bar. So the combination of mass production and robotaxi is super important for us. We have also a third vertical where we provide a redundancy to both of these systems. We call it air-to-road. air-to-road essentially is the capability of leveraging the aerial modality, aerial images, whether these are satellite or drone images acquired offline. And letting the car basically this kind of unfair advantage of knowing its surroundings without using expensive HD maps. So these are the three vectors, which all of them converge to the same mission, enabling every car to become autonomous without investing in very high compute in high-end sensors like LiDAR, but basically using basic five, seven cameras, available compute. And release the driver from supervision. And you always ⁓ mention also that you're software only and chip agnostic ⁓ is kind of your edge as well. So ⁓ what does staying out of silicon let you do and what does it cost you in control, maybe?

Daniel Abreu Marques: And you always ⁓ mention also that you're software only and chip agnostic, ⁓ and it's kind of your edge as well. So ⁓ what does staying out of silicon let you do and what does it cost you in control, maybe?

igal: No, the the philosophy of auto brands is basically to do the minimally needed portion of ⁓ the value chain that is you know is required to to reach the vision of autonomy for every vehicle. And so there is ⁓ plenty of ⁓ amazing hardware players, NVIDIA is at the top. They have ⁓ the best system on the chips and and a very good ecosystem, which we are benefiting from. From the hardware itself to simulation environment, Cosmos Alp and Alpasim, to also world models, which are benefiting our agents. there are many other players. ⁓ if the OEM prefers to use other SOC, the technology can be ported as well. So we definitely stay as hardware agnostic. I would even say ecosystem world model agnostic player, as an agentic layer on top of all these ecosystems that are being developed for the broader. scope of physical AI. And you also mentioned that you already have production programs at level two plus ⁓ plus plus. ⁓ the new headlines are now level four robotaxis. ⁓ guess that the the biggest part of your company is the is the ADAS business ⁓ now the level four business is is coming in addition. ⁓ and how do you balance the two and how do you envision also the the ⁓ to develop over time?

Daniel Abreu Marques: You also mentioned that you already have production programs at level two plus or plus plus. and the new headlines are now level four robotaxis. I guess that the the biggest part of your company is the is the ADAS business and now the level four business is is coming in addition. ⁓ and how do you balance the two and how do you envision also the the shift to develop over time?

igal: I think we see a convergence between the two. When we started the mass productions, we started from the most high volume segments, which is the front camera for safety and regulation. today the new programs that we have secured are focusing already on what is called navigation on pilot, essentially Tesla-like functionality, the ability to drive in and urban scenarios. So this a system that should enable the driver with the full functionality of autonomous driving in a regular car. But the driver, of course, is still supervises the system. The Robotaxi on the other hand, allows full L4, right? So there is no supervision obviously, there is even no driver in this on in the seat. But it's ⁓ it's ⁓ doesn't have to meet day one, those performance in terms of cost and in terms of covering the whole world. And we see that during the next year those two vectors, with the third one being air-to-road which is a redundancy, converging essentially having a single product which is ⁓ autonomous driving affordable for every car.

Daniel Abreu Marques: sense and what's also super interesting ⁓ with auto brains is your investor list which could be doubled also as your customer list more or less so you have BMW Toyota Continental Vinfast Magna Knorr Bremse so a powerful list ⁓ and how do you manage being funded by companies you also sell to?

igal: Makes sense and what's also super interesting ⁓ with auto brains is your investor list, which could be doubled also as your customer list, more or less. So you have BMW, Toyota, Continental, Winfast, Magna, Knorrbremse, so a powerful list. ⁓ and how do you manage being funded by companies you also sell to? So this are you know a very good synergy with the several many of the investors, they're also like I mentioned, their partners and customers. ⁓ every such partner adds a lot of validation on the technology. As you mentioned, this is a contrarian approach, right? So we we believe this is a disruptive technology. It doesn't mean that ⁓ it's not building on top of everything that was developed so far, but it's a new layer, a new evolution of the technology. And so it's ⁓ really in the automotive world, which is extremely conservative is super important. And part our fundraising process was really focused on find partners and investors that would be the leading players in automotive world, whether it's from tier one perspective, OEM perspective, silicon perspective, that would you know ⁓ let pass through this scrutiny, testing, you know, comprehensive due diligence, reviews, tests in the car, huge data sets. And this of course gives us a lot of credibility when we go to new customers. Yeah, absolutely makes sense. And and where's the line ⁓ between what Autobrains builds in-house and what you hand to partners? So what would you never outsource? we believe that there are few pillars that are very unique to us, right? So the technology is based ⁓ on this agentic AI approach.

Daniel Abreu Marques: Do it to make sense and and where's the line ⁓ between what Autobrains builds in-house and what you hand to partners. So what would you never outsource?

igal: It is based on the air-to-road that I mentioned that provides kind of a copilot for every car with the aerial dimensionality. There is a very unique IP ⁓ in we call common sense, the ability to learn by thinking and the ability to think while driving. And there a fourth pillar that we're going to introduce soon. We believe those unique ⁓ unique verticals ⁓ add a of value to the OEMs, but there are a lot of common developments and world model is a good example, right? So everyone is is ⁓ basically there is a a race to generate the best world models and so There we prefer to be neutral approach. We of course provide our capability as well, but if the OEM or any other customer has its own world model, in many cases it can even be beyond the world of autonomy, autonomous vehicles, it can focus on broader physical AI area. So we can be agnostic and we can leverage from the assets of our partners.

Daniel Abreu Marques: Makes sense and ⁓ now I want to dive into the news that that we already covered a bit. So you had two announcements ⁓ both with NVIDIA, so maybe let's take them one at a time and start with Munich. So in your own words, what did you actually announce and what is this Robotaxi program about?

igal: Makes sense and ⁓ now I want to dive into the news ⁓ that we already covered a bit. So you had two announcements ⁓ both with NVIDIA, so maybe let's take them one at a time and start with Munich. So in your own words, what did you actually announce and what is this Robotaxi program about? So the announcement is about launching Robotaxis in Munich based on NVIDIA architecture and NVIDIA ecosystem jointly with Uber in part Partnership with Uber and the Ride Hailing Solution and the whole ecosystem of distribution and operation, where the stack will come from Autobrains, the autonomous stack. this program OEM agnostic. We have ⁓ currently few OEMs, but we we keep it agnostic so every OEM can join it. We believe that Munich is ⁓ in particular very interesting as the regulation, the safety bar is the highest there. So once you prove something in Munich. The opportunity to scale is very fast to any other city. the release of this PR, we got a lot of interest from OEMs and other players to join this program. And have a very, very aggressive timeline already this year and early next year with some milestones that we'll be able to announce as we meet them. ⁓ we're very excited about that. There is already the first car running ⁓ in of course with a test driver in in the mode that you can feel the results. We launched it three days after we received the car and the permission. And then the third day of driving, we already experienced very smooth, you know, zero intervention behavior, with very good feedback and very complex environment in ⁓ Munich, speed urban environment, very dense traffic, very narrow streets, a lot of construction sites, a lot of cyclists coming on ⁓ crosswalks and so on, and we we are very confident about the system.

Daniel Abreu Marques: Yeah, sounds really good and really good to see this movement in Germany for me as a German. So great that you do that there. ⁓ and for someone who's hearing ⁓ OEM agnostic for the first time, so what does it actually mean in the Munich program and why is this the right model instead of going deep with one car maker?

igal: Yeah, sounds really good and really good to see this movement in Germany for me as a German. So great that you do that there. ⁓ and for someone who's hearing OEM agnostic for the first time, so what does it actually mean in the Munich program and why is this the right model instead of going deep with one car maker? You know, since our approach eventually is to get to every car, we want this capability to to be exposed to any OEM that is interested to insight into this agentic AI technology. When we deploy those Robotaxi they look like normal cars, right? So it's not that you will see a few lidars and a lot of GPUs in the trunk and some geofenced HD maps. It looks like a really normal car, affordable for every car. And so we believe that keeping it open to every OEM is very important because OEMs are interested in massive scale. And for them it's an opportunity to experience the technology firsthand.

Daniel Abreu Marques: And ⁓ as we talked about before, you also always highlight that you're chip-agnostic ⁓ and now in Munich you're partnered with NVIDIA with a drive hyperion platform for these level four programs. And is this then still ⁓ agnostic or is this exclusive ⁓ with NVIDIA or is this still an option among several chip players?

igal: And ⁓ as we talked about before, you also always highlight that you're chip agnostic. ⁓ and now in Munich you're partnered with NVIDIA with a drive Iperion platform for these level four programs. And is this then still ⁓ agnostic or is this exclusive ⁓ with NVIDIA or is this still an option among several chip players? It's agnostic. ⁓ definitely we are agnostic and for for mass production ⁓ programs we we also use different SOCs, but ⁓ we We believe that when we think about scalability, NVIDIA is a very strong platform. It emerges as the leading platform. It goes way beyond GPUs, but and when you think about scalability, scaling to full autonomy, I think it's a very good solution. And what does Nvidia provide, autobrains on top of GPUs?

Daniel Abreu Marques: And and what does NVIDIA provide autobrains on top of the GPUs?

igal: So NVIDIA provides us ⁓ with a lot of support, both the the GPUs, the ⁓ the ecosystem that ⁓ you know there is a ⁓ basically a huge world model, ⁓ autonomous driving model, end-to-end model of ⁓ billion parameters growing to 100 billion parameters called Alpamayo, ⁓ the one that was launched a few months ago, and we have access ⁓ to this ⁓ model and it shows an amazing performance. And ⁓ using our agentic AI layer, we can basically distill this model into very tiny agents that can run in the car. it Also provides us with a lot of simulation environments like Alpasim and Cosmos, a lot of tools related to the physical AI world, and the packages related to safety. So there is a lot of ecosystem built in Nvidia, a lot of platform capabilities that we can leverage. ⁓

Daniel Abreu Marques: And I I think you probably cannot name an OEM yet, otherwise you would have publicly announced, but without naming names, ⁓ what of automaker does this program need and what are they really signing up for?

igal: And I I think you probably cannot name an OM yet, otherwise you would have publicly announced, but without naming names, what kind of automaker does this program need and what are they really signing up for? So we're talking about global OEMs that's important to say, that when they look at this program, they always think far beyond. They think about mass production. So they want to see the cars driving in Munich, but how this would translate to making drive anywhere in Europe, you know, far ahead of central cities and mountains and countryside, anywhere. And the work with OEMs once have the Uber support in start in terms of demand and ⁓ distribution and operation, and we have the NVIDIA hardware stack, Hyperion, and we have our software stack. I think the work with OEM becomes very smooth and ⁓ very low friction ⁓ and them access to this technology. So I think we we have created a very unique approach to OEMs.

Daniel Abreu Marques: Super interesting. And maybe now go to the second announcement. So tell us more about Southeast Asia. So what did you and Vinfast announce and what's the goal there?

igal: It's super interesting. And maybe now go to the second announcement. So tell us more about Southeast Asia. So ⁓ there? So we are working with Vinfast ⁓ on ⁓ the full spectrum of ADAS even before announcement. important to say we are providing to them safety and ⁓ autonomous driving products from entry level. As mentioned, the braking and adaptive cruise control and sign recognition and traffic light recognition and lane keeping and many and many other features at a very, very affordable cost. That's very important for this geography. beyond that, it has to be localized. So there is a very high challenge to localize those capabilities to South Asia environment with very chaotic driving patterns, a lot of two-wheelers, a very different structure of the streets. And we are covering today the full spectrum from L2 all the way to L two plus plus. And the recent announcement basically also ⁓ this to Robotaxi When we say robotaxi, we think about very affordable autonomous vehicles. This is also in partnership with NVIDIA, we're from based on the same Considerations that described in the Uber case, giving us the ecosystem, a very good solution hardware-wise. And VinFast is a very strong player in South Asia, as you know. They're scaling in India, in Indonesia. They are beyond being just a pure OEM, they're an ecosystem player with the whole charging infrastructure and their robot tech and their taxi company that has about 30,000 cars. they are ⁓ very significantly volumes. ⁓ And this is a very good opportunity for us. And Vinfast is, I mentioned before, also one of your investors. So, how is that partnership structured differently from Munich? And maybe what does Vinfast own and what does Autobrains own? it's very similar. So Vinfast is a is is an investor, ⁓ of a very large set of leading investors that we have in the company. so we have both ⁓ and the

Daniel Abreu Marques: And Vinfast is I mentioned before also one of your investors. So how is that partnership structured differently from Munich? And maybe what does Vinfast own and what does Autobrains own?

igal: We have them both as investors but also very strong partner. general, we are working with OEMs in a very open, collaborative model, right? So the idea is not just to sell OEMs black boxes, but to empower them with this ⁓ AI infrastructure where they can fine-tune, localize their products, where they can support multiple car models and brands and so on. ⁓ And ⁓ we have a very open collaboration, a very good collaboration with Vinfast Sounds good. And when we look at Munich and Southeast Asia, ⁓ I think these are almost opposite worlds ⁓ probably in terms of driving behavior and ⁓

Daniel Abreu Marques: also maybe driving so dense and unpredictable traffic and other points. ⁓ but you have the same agentic system in both. So what actually has to change ⁓ those deployments?

igal: Also maybe driving environments or dense and unpredictable traffic and other points. ⁓ but you have the same agentic system and both. So what actually has to change between ⁓ those deployments? Yeah, so so the the Reason you know basically the fact that we have this agentic AI approach allows us to scale faster and separate a lot the problems. So imagine that you have in South Asia an agent that is ⁓ focused on passing junctions, right? And then we get to Munich and you have a very different behavior in junctions. So on one side, of course, we can benefit from all the Knowledge that the agent learned in South Asia to Munich, but we don't have to stretch the system to address both environments and very different environments and the same system, basically kind of getting to average performance in each of them. So we have a special agent here and a special agent there, and there is a lot of knowledge and transfer learning going between those agents. But on a more high level, yeah, so it's it's both of them are challenging. The South Asia, as I mentioned, is very chaotic. you analyze kind of the traffic patterns, it feels more like a diffusion. There is no very clear laws and can be always surprises. In Munich, the other hand, it's urban environment, especially when you focus on narrow streets, two-way narrow streets with construction sites, cyclists, and so on. And you cannot be too slow. That's one of the things we've learned very fast on the third day. You must drive at very close to maximum speed, of course, very cautiously when you can. But if you are too conservative, that's not good enough. So you have these both ends. You have to be super, super safe, but you cannot be too conservative. The car has to be assertive, otherwise, it will be you know stuck forever in different merge scenarios. And manage to address both.

Daniel Abreu Marques: And from a from a sensors perspective perspective would be the s would it be the same sensor setup in every region, or do you also have to have different sensors for different regions?

igal: And from an from a census perspective. Perspective would be the would it be the same sensor setup in every region, or do you also have to have different sensors for different regions? No, no, our approach is the same, exactly the same system. Exactly the same system, hardware wise and sensor wise. The philosophy is vision the ⁓ based, right? So vision only should solve the whole problem. If there is an interest for other sensors, for example, from OEM perspective or any other customer perspective, we might add additional sensors, but they would only serve as a Redundancy. Our mission as being an AI company and reproducing the human driver behavior and scaling it from there is to be based on vision only sensors. And ⁓ maybe ⁓ frame it like this: Munich is also a bit ⁓ statement about Europe. ⁓ we can now let's zoom out to the whole continent. ⁓ ⁓

Daniel Abreu Marques: ⁓ maybe let's frame it like this. Munich is also a bit a a statement about Europe. So we can now let's zoom out to the whole continent ⁓ and maybe ⁓ re reiterate again. So why Munich specifically? You framed it, it's a tough regulation, also complex roads and it's a feature, not a bug, more or less. ⁓ could you make the case why starting from somewhere hard?

igal: Re-reiter it again. So, why Munich specifically you framed it? It's a tough regulation, also complex roads, and it's a feature, not a bug, more or less. could you make the case? Why starting from somewhere hard? I think you know, if you want to scale, you need to start from the hardest place, right? You want to show that in in the area where you have the concentration of the most difficult scenarios, when you have this combination of high performance. But also very high safety and regulation requirements. If you perform here, obviously you can scale to a simpler problem or to a similar problem. So that's the reason we focused on these two geographies. We still believe South Asia is one of the most difficult places in terms of let's call it edge cases, because the edge cases in South Asia are the main cases, the regular cases. In Munich, on the other side, it's a you know probably typical European challenging city. But also with the highest regulation and safety requirements. And you know, we have the benefit of having a site here, R&D center, essentially our headquarters. having presence here definitely helps with the connection to the leading automotive players, both local and global. So it was natural for us.

Daniel Abreu Marques: Can you give us ⁓ so some context on how big your R and D center is currently in Munich? I think it I don't know it's how how long you're been there, but maybe some context on on the the size.

igal: And can you give us ⁓ some context on how big your R and D centers currently in Munich? I think it yeah, I don't know it's how how long you have been there, but maybe some context on on the size? We have a bit over 20 people here. The whole company is just over hundred employees. So it's the second you know largest presence for us. We still try to to be a very small company. As we say, we are a hundred employees and two thousand agents, coding agents. but yeah, 20 people and growing in unique. ⁓ interesting. And ⁓ is ⁓ usually described as behind the US and China on AV deployment. So ⁓ do ⁓ buy framing or is Europe playing maybe a different game? So what are your thoughts on Europe? I think Europe is very strategic to us. Europe is more difficult than US, even if ⁓ you start you know the streets, right? There are very

Daniel Abreu Marques: Super interesting. And Europe is also usually described as behind the US and China on A V deployment. So do you buy that framing or is Europe playing maybe a different game? So what are your thoughts on Europe?

igal: wide streets ⁓ much less dense traffic at least where the robotaxi were deployed so far. obviously US is a very important market, but the focus for us in the next 12 months is going to be Europe and South Asia. That's more than enough for for this time frame. And I Europe ⁓ also ⁓ leads many in terms of let's call it kind of automotive conservatism. So again, if you if you make it here, scaling to US would be easier. And and how is your read on the regulatory part? So what ⁓ specifically has to happen in Germany and EU for Munich to move from pilot to an actual paid service? We have a roadmap that has several key milestones. one of the key milestones is to start at very aggressive timeline commercial drives.

Daniel Abreu Marques: And and how is your read on the regulatory path? So what ⁓ specifically has to happen in Germany and EU for Munich to move from pilot to an actual paid service?

igal: So that's what we are working on now to release ⁓ to support the cars for commercial drives where we can actually take passengers in Uber Network. This is still going to happen with the test driver. this is already a very significant milestone because the car should basically function as autonomous car with no interventions, very high comfort. So that's the first milestone. The second milestone is then providing the service for a certain period of time and collecting enough data where we can prove that we have less than one intervention ⁓ ⁓ than 50,000 hours. So that's a superhuman ⁓ ⁓ As we say, it should be still human-like, right? So even those mistakes and a huge data set, which are much less than frequent, should be logical. So it should be logical for the human to understand what happened. But once we reach this level, we can take the driver out. And then basically we will be the first company to operate full autonomous cars in Europe.

Daniel Abreu Marques: And one one thing that you always hear in terms of European deployments is also type approval. So you're working towards a type approval with OEMs

igal: And one one thing that you always hear in terms of European deployments is also type approval. So you're working towards and type approval with TOMs. Yes, we we are we have a a team, a safety team that work very closely with the local players here, both the authorities and the local companies that are experts in the specific processes. So that's ⁓ part of the process that we're doing. Sounds good. And

Daniel Abreu Marques: Sounds good. And do you already have an a roadmap of European cities that you want to tackle after Germany, or is it currently still f only focused on Germany and everything what's ⁓ comes after that is ⁓ s ⁓ part?

igal: have a plan. We have a plan. We plan to scale to 20 cities, but we we still cannot release it. Okay So ⁓ maybe ⁓ one more question to close that looks more ahead. So five years out. ⁓

Daniel Abreu Marques: Okay. So maybe one more question to close that looks more ahead. So five years out, what does a realistic European Robotaxi map look like to you and where does Autobrains sit on it?

igal: What does a realistic European RoboTeximap look like to you and where does autobrains sit on it? I think in five years there will be two big trends, right? The robotaxi will s will scale from thousands to tens of thousands and potentially will cross hundred thousand robotaxis overall, maybe even hundred robo in Europe. So all the major cities will have Robotaxis. But there will be also a very important additional trend. There will be millions of cars ⁓ will have eyes-of functionality, at least for major periods of time. So you will be able to drive your own car while you're reading emails, working, watching ⁓ you know video. and not having the eyes all the time on the road. And those those two trends somewhere actually around five years from today will converge. So there will be no such it will become a bit blurry the distinction between a private car and a robotaxi car. The robotaxi car will look like more like a private car and the private car will become more and more autonomous.

Daniel Abreu Marques: It's super interesting, especially the eyes-off part. So it's then more or less a level three system as we call it. But oftentimes there's also some critique for to the level three system because you you think you're buying an autonomous system, but you still have to be you need to be able to grab the attention or the the control of the vehicle in a a seconds amount.

igal: That's super interesting. ⁓ especially the ISOF part. So it's then more or less a level three system as as we call it. ⁓ but oftentimes there's also some critique to for to the level three system because you you think you're buying an autonomous system, but you still have to be you need to be able to grab the attention or the the control of the vehicle in a ⁓ a seconds amount. So ⁓ how do you see that? Is it still really like a compelling value for the customer?

Daniel Abreu Marques: So ⁓ how do you see that? Is it still really like a compelling value for the customers so that they want to pay extra for that?

igal: So that they want to pay extra for that? I think that's the most compelling value for the customer. Imagine if you would have a car where you can even take brakes from driving, right? Of course, the level three systems have to be implemented properly. So it cannot surprise you and ask, you know, take the control in two seconds or in 10 seconds. So the implementation is super important. I think got a bad reputation ⁓ it was very expensive. It was only luxury cars. It was extremely limited and it was relatively poor, right? So you really couldn't, you know, there was no system you could take a break from driving in urban scenario. there not such a big difference from in terms of functionality from just you know adaptive cruise control on highway. And that's the reason I think that L3 got a relatively bad reputation. There are another reason that there is a bit of a friction between what the driver needs and what the OEMs is ready to make, right? So the driver definitely needs eyes off, right? We all want to to be able to do things while we are driving and not kind of sit with the hands off and have the eyes on the road, kind of certain punishment. ⁓ but ⁓ the OEMs need to take the liability and the responsibility over their autonomous driving systems once the driver takes the eyes off. And that's a very big kind of step change, right? and that's where the system should show a completely different level of performance.

Daniel Abreu Marques: And ⁓ do you think that is it's no problem to ⁓ get the situational awareness? For example, if you're traveling on a highway and you're doing other stuff, may you watching a movie on a tablet or something like this, and then the the car prompts you to take over and ⁓ you need to immediately step over. So do you think there's a kind of a problem and you need to kind of have training for the drivers of L3 cars or something like that? So how to

igal: And ⁓ do you think that is it's no problem to ⁓ get the situational awareness? For example, if you're driving on a highway and you're doing other stuff, ⁓ you're watching a movie on a tablet or something like this, and then the the car prompts you to take over and ⁓ you need to immediately step over. So do you think there's a kind of a problem in you need to kind of have training for the drivers of ⁓ three cars or something like that? I think what they describe I think what they describe what it is exactly what happened with level three. And it's a terrible experience, right? If if you if you kind of have to be able to take the control from the car in a fraction of ⁓ in a few seconds, in 10 seconds even, it keeps you really kind of still anxious and you cannot relax and you cannot forget about this you know responsibility to take over the control. When I talk about level three, I'm talking about a very planned transition of control. So for example, when you plan your drive from home to work. The car can tell you, okay, these 10 minutes I'm going to drive myself, but I'm going to alert you three minutes ahead of time in a certain junction or in a certain ⁓ part of the road that I will need you to take the control. And so you know, and you can plan ahead of time, and it's very clear the threshold between the driver responsibility and the car responsibility is super clear, and it cannot happen in a matter of seconds, this transition.

Daniel Abreu Marques: That sounds a lot better from from the value perspective for sure.

igal: That that sounds a lot better from from the value perspective for sure.

Daniel Abreu Marques: ⁓ so so we're at the end. Igal, I thank you really much for taking the time today and shed some light on on the news of autobrains and I'm really excited to see developments in Munich, especially. ⁓ love to visit you when I'm in Munich again, ⁓ to see what you're doing over there. And ⁓ thanks Igal. So happy to get you on the show today.

igal: ⁓ so so we're at the end. Igal I thank you really much for taking the time today and shed some light on the news of auto brains and I'm really excited to see your developments in Munich, especially. would love to visit you ⁓ when I'm in Munich again ⁓ to see what you're doing over there. And thanks again, Igal. So happy to had you on the