Episode 103 – Cracking The Code Of Automotive Radar: An Inside Look With Arbe’s Chief Radar Officer

Autonomous Driving Technology
In this episode, Bill Brock interviewed Gonen Barkan, Chief Radar Officer at Arbe. Gonen shared his extensive experience in the world of automotive radar. He offered a deep dive into the challenges and breakthroughs in building reliable, high-resolution radar systems for autonomous vehicles.

They explored the complexities of radar technology and how it differs from traditional vision-based systems. They also highlighted Arbe’s innovative approach to “imaging radar” that provides a rich, multidimensional data stream to power advanced autonomous driving features. Gonen discusses the importance of radar resilience in adverse weather conditions and how his team tackled unique challenges like detecting objects buried under snow.

They also discussed the broader landscape of the automotive industry. Barkan reflected on his time at GM and the struggles traditional automakers face in adapting to the rapid technological shifts driven by companies like Tesla. He shared his perspective on the role of software and the need for a shift in mindset to embrace autonomous driving capabilities.

About the Show

“Over the Air” is a podcast that features unfiltered conversations with executives about their IoT journeys. The mistakes they made, the lessons they learned, and what they wished they’d known when they started. In each episode, our expert guests make bold predictions, reveal unpopular opinions, and share stories you won’t hear anywhere else.

Buyers Guide – How to Choose the Right IoT Company: Very

Subscribe here

YouTube

Apple Podcasts 

Spotify 

Transcript

Hello, everybody. My name is Bill Brock, CEO of Peridio, and I’m host today with Over the Air podcast. And today we have a very exciting show going on Barkin from Arbe.

He’s the chief radar officer and has a very interesting history and story to share about building technologies for autonomous vehicles. So welcome to the show. Thanks for joining.

Thank you very much. Great being here. All right.

So for those who are not familiar with you or your company, would you mind just giving us a little bit of background in history and what you guys do? Yeah, so Arbe is a radar company that is super focused on autonomous driving and autonomous driving features, you know, bringing the true safety to everyone and everywhere. Its core technology is chipset, the silicon with the highest resolution in the automotive world, about 10 times what’s available in the market. Our plan is to be like the breakthrough in vehicle safety.

Very cool. And chief radar officer, you’re the first one of those that I’ve ever met. So tell me, tell me a little bit about what does it mean to be a chief radar officer? Yeah, so for those of you who, you know, who are not in the radar business, radar is actually doing the radar is way beyond just, you know, silicon and chips and RF.

There’s a lot of algorithms and software and how those you need to deal with the real world, right? Well, it’s not just software. You need to figure out what happens in the rain. What happens when the vehicle is bouncing? What happened when your bumper, you know, we’re officially get bumped.

So how do you deal with all those things and get it into cost? And this is why kind of my expertise in is getting the right, you know, the right technology with the right silicon with the right software to the right customers and right time for the right features. So it’s kind of all around. This is why we needed a special title because it’s kind of touch everything from the customer down to the silicon level.

What’s really fascinating industry and the more and more I learn about it, it’s so hard on the hardware and the software front. And where do you, where is your like primary background? Are you more of a hardware guy? Are you an algorithm guy? What’s been your background and focus? And yeah, yeah, so it’s never one of the other. I’d say that I’m more of a digital guy and algorithm guy and how do you handle the real complex world and, you know, figure the right, you know, math into it and write implementation into it.

But definitely tons of hardware. Eventually hardware is hard. This is why it’s called hardware.

And you have to deal with RF and especially on the high frequencies. It’s like voodoo. So you have to be all around and first and foremost, you have to be customer-centric.

You have to understand what the customer needs and not what the technology is able to do. You need to be able to do this translation. And that’s I think the hardest part for engineers.

Yeah, and so what is talk to me and it’s going to be hard not to get too technical here, but at a high level. Tell me a little bit more about the stack and how these systems operate. The big challenge here is that when we see the world, you know, vision, which is very trivial to all of us, but, you know, have you seen the predator, you know, the classic, right? So there’s a scene when, you know, his switches filters.

Yeah, he goes to the infrared. It’s kind of like this, when you try to see the world through RF not for vision. It’s extremely hard for us as humans to understand, especially as a picture is not bi-dimensional.

It’s not XY. It’s six-dimensional. Okay, so we don’t have even have the vocabulary to understand what we’re seeing and not even to visualize it.

So radar is kind of composed of all of these. First, it’s a physical world. So you send very, you know, high-frequency signals into the world and understand, you know, how materials impact them, how the environment impacts them, how weather impacts it, and then you want to do it very low-cost because you want it to be mass production.

So you need to understand the trade-off between get high fidelity information from the environment and getting low cost, which always is counterproductive, like you’re trying to see the world for a 0.3 megapixel camera, not for a high-definition camera, and still figure out what’s happening in the world. And then you need the processing and you also want it to be cheap. So you need to find very unique algorithms and ways to implement them in a way that is not brute force, that is very sophisticated.

So it’s not sufficient to know the solution. You need to find the right solution and that you can implement at cost and fast. And then you need to understand very well what the customer needs, which may be a perception system, which may be even some some visualize it or fusion systems or decision systems.

So how do they need to get the information? So all of this is complexity at multiple levels that you need kind of figure out and also control the cost all the time. Talk me through maybe an example use case that folks could really relate to. Is this, are you guys working in consumer automotive or is it more forward self-driving technologies? Like where’s the the primary customer base right now? Where are you focusing? Yeah, so for Arbus specifically, the target is consumer, what we call consumer autonomy.

So L2 plus and above. And so what does that mean? L2 plus and what is L2? So the right taxonomy and thanks Mobileye for helping the industry understand this. So L2 plus is the first time actually it’s L2 but practically L2 plus first time you can take your hands off the wheel.

Okay, still need to supervise. You need to be capacitated. You need to supervise and you’re responsible.

Right, but it’s not just alerts. It’s the point where the vehicle starts to make decisions. Okay, and the lowest end possible that I can think of which is not even L2 is A-E-B emergency braking.

A ton of emergency braking. We, I don’t know if you’ve seen the latest analysis by IHS and NHTSA, it’s the performance are really really really bad even at low speeds. Once daylight or dim light, it simply doesn’t work.

So from these elements that actually saves lives up to L3 when you you’re allowed not to supervise the driving but you have to be capacitated to take over. But last but not least is L4 which is robotaxis where you don’t need the capacitated driver in the car. So all of these from, you know, real true A-E-B, true safety A-E-B, right, up to robotaxis.

So that’s our domain where you really need reliable sensing to be able to get the safety levels. Cool, and so can you speak to any current customers that you have right now? Any products that are in the field that customers are not like the listeners could relate to? So, although it’s a radio company, eventually the products we’re selling is chipsets. We’re a classical two company in the supply chain for radars.

We have our customers are the big two ones. We have announced three two ones. We have Magna, which is our western tier one and two Chinese tier ones.

We Fu and the Rain and we definitely working on more and they are the ones who are actually selling to the OEMs. So you’re focused on the chipset and then you would then bring in off-the-shelf radar sensors. Is that correct? So we developed the chipset but we developed our own reference design for the radar to prove out that it can outperform all the other is radars in the world.

But we do not sell this radar. We provide this reference design to our two ones. But our two ones are radar two ones.

They’re building tens of millions of radars a year. So they take our chipset. They take our reference design and they build their own radar and they are the one who’s bidding to the OEMs.

We’re not selling the radars. Very cool. So sounds like there’s a little bit difference with this new radar technology that you guys are developing.

What is really the importance of an imaging radar and what is different about this approach? Yes. So first and foremost, we need to understand what the term imaging radar really is or at least how do we call it? Okay, sometimes we call it even a perception radar, not an imaging radar. You may encounter terms like 3D radar, 4D radar.

I think that there are two key differences. Imaging radar is a radar that is inherently what we call unambiguous, meaning that it provides reliable data at all situations, at high resolution, without relying on pre-knowledge of either the maneuver or the feature or how it’s going to be used or the environment it’s going to be used. Okay, in this sense, just like a camera, right, when you design a camera, you don’t say, okay, it’s only going to work on the day at Moab Desert, right? You say it’s going to image the world at any situations.

Now all of the radars today, the legacy radars in the market, like 99.9% of the radars, they utilize the fact that it is used for very basic, they’re not used for imaging. They’re used for very simple tasks, like adaptive cruise control. You only need to track a car in front of you.

So there’s a lot of pre-assumptions in the design that allows you to use lower end hardware, okay, with less resolution, less processing power, less reliability, not as rich data, because you only need to do very simple tasks. Once you try to image the world like a camera, but with an RF device, with a radar, then you cannot make those assumptions. So you need much more channels, or if you want the equivalent of pixels, you need much more processing power, you need to remove all those pre-assumptions on the use and produce a much richer data.

Now this is, on the current technology, it’s very expensive. Arbus technology allows you to get 10 times or 20 times more resolution and processing power for about the same cost of the best in class radars today. And that’s the key difference.

So what is it? And I think we touched on this a little bit earlier, but Sam, and I’m not a self-driving engineer, so this might just be a total rat’s nest of complexity that we’ll have to quickly back out of. But what are like the signals and interfaces that you provide to the engineering team that you are using these sensors? They say like, hey, there’s an object, 20 feet out, coming this way, how does somebody actually go in and work with the signals and inputs from your system? So I’ll use again the analogy from vision, right? When you take a camera, you take all the information from all the pixels, right? You’re not asking the camera to, I mean, there is of course, perception algorithms and software stacks that try to understand the image from the image, what is happening in the image, right? But the sensor provides image of the entire wall. It doesn’t decide what to image and what not to image.

The sensor role is to image the world in the highest fidelity possible. So the perception engines and the scene understanding engines will be able to make out out of the sensor image what happening in the environment. So unlike all the legacy radars, what imaging readers are doing, they’re providing what we call a point cloud, which is if you want very simple to pixels.

But when you get information from a camera, you get x coordinates, y coordinates and usually in RGB, right? Yeah. So from a radar, you get x, y, z. So three-dimensional because it measures depth. It measures the velocity of objects or the Doppler and it provides you also the magnitude of the return.

Okay, and you get this for anything that has any return in the environment. So you get a very dense point clouds that are describing the world and four dimensions, x, y, actually five, x, y, z Doppler and power. Okay, so that’s what you get.

So if five-dimensional point cloud of tens of thousands of points that you need to figure out, you know, what does it be? Sure. And so then there’s the the interpretation and training of the models and everything, but you’re a very high-quality sensor input to these self-driving systems. Exactly.

Cool. Very interesting. What are the use cases that you’re most excited about? And I have here a note, autonomous trucking that seems to be an emerging industry for you guys.

So what is future-looking the opportunity for you? Yeah, so undoubtedly, Robotext will get there at some point. And I think we see more and more higher autonomy features on retail vehicles. But I think autonomous trucking, I think is the one industry that where the value, the business value for autonomy is the most evident.

Either from safety perspective, either from ability to operate around the clock, even from if you can drive even faster. Now, what’s special about trucks that is very different from customer vehicles is that you don’t have so much design limitations because eventually in radar, the bigger the aperture, the bigger the size you get, you get better resolution and better performance. It’s much easier to package those radars on big trucks.

Right. They’re not as worried about the form factor. Yeah, the cost sensitivity is not on fleets and on big trucks is not similar to customer cars.

And first and foremost, because of their mass and speed, they need much longer detection ranges and much better reliability than customer vehicles. And this is where, this is where I think the benefit of high-end radars will shine first before any other platform. And the business value is very clear.

So the business case is very clear. Right. So these trucks are out there performing a job.

You can help folks drive safer, longer, be more efficient with their work and their performance. And also, I think the biggest value I think is that it will allow you to operate at any weather, any lighting conditions, fog, dust, which today for the vision-centric systems, not only they do not have the redundancy to get to safety level, they simply fail on weather, on lighting conditions, on glare to try to drive against the sunset. So there’s a lot of a very big part of the day or that geographically they’re simply ineffective.

How do you guys handle this? How do you compete here? Like, what has been your real advantage in handling all these different scenarios? Yeah, so it all comes down to physics. We work, we’re not a vision system. We work in a completely different frequencies, like 80 gigahertz, 77 gigahertz, which are completely unaffected by those factors.

Definitely not by lighting. It’s, you know, radar are completely agnostic to sunlight or daylight or whatever. It’s completely different frequencies.

So day and night is exactly the same performance. And dust and fog has very little impact. It does have an impact or heavy rain, but definitely much more resilient to adverse weather than any vision system.

What about, like, I used to live in Wyoming for a blip, like two years I was out there and we’d have these, they call them whiteouts, where the, it was not snowing, but the wind just kicks up snow and all of a sudden it’s just getting blown across the freeway. You can’t, does it work in that scenario as well? It’s horrible. I think it was never tested to this extent.

It’s definitely less, as I said, it’s always have an impact, but it’s dramatically less severe than with vision. And I can tell you a story, you know, for one of my first, I worked at General Motors for a lot of years and I developed radars for crews in San Francisco for the robotaxis and we went to do some field tests. So it was snowing in Detroit, January and what we’re driving, everything is white, like two meters of snow.

And suddenly in the radar, I see like every two meters I see a blip and I look outside and I see nothing, say, what the hell is it? I have no idea. And so we stopped the car and my engineer, you know, went down, went out of the car and I was guiding him through the radar to go to this blip. I cannot see anything.

I told him, just clear the snow. He’s clearing the snow and you know, those little poles that you put on Santa Rose to know where the broad edge is. Oh, yeah.

The little metal poles. So as big as I saw them underneath the snow. Unbelievable.

So that is wild. So it’s really a matter of understanding that this is not vision. Okay, we’re not used to it.

It’s very hard. And this is, by the way, one of my biggest struggles, you know, when I was, I was demoing to GM officials, like, I know, number three, number two is in GM and in cruise every three months. And these are not radio guys.

They’re vision guys. All of us are vision guys, right? That’s what we know. So you’re trying to show radar output to people that have never seen how they look like, like a five-dimensional point cloud.

So a lot of work is on how to visualize it to someone who’s not in this field to make sense. That’s very cool. I don’t understand it either.

That is truly mind-bending. Same through snow. That’s awesome.

So you mentioned working previously with GM. You’ve got a very interesting background. So could you tell us just a little bit about the path to chief radar officer? What was that maybe like the last 10 years? What did that look like before getting into current role? Yeah, well, chief radar officer is kind of an arborist terminology.

I actually started for I think I was 15, 16 years in the startup world. I was working in communication. I actually did I think back in 2001.

I think before Bluetooth, before Wi-Fi, before anything we did power line communication that could go 10 times the Ethernet speed between all your outlets at home. But eventually Wi-Fi and Bluetooth came in and so we had a lot of people who sold the market. So I did from algorithms to chip.

I taped out six chips for this domain. Some of them are still being sold today in millions. So I’m proud of that.

And sometime 2013, a friend of mine who worked at GM is a researcher on radar. He told me let’s just spin off from GM and do imaging radars for Thomas Riven. It looks cool.

Let’s be the pioneers. And you know GM, Israel General Manager Gil Golan at that time, he had a vision. He says come do it in GM.

Tell me how many people you need, how much time do you need and what do you need? You have 10 months to do the first imaging radar with the first CMOS chip in the world that can support it from TI. And that’s what I did. Very cool.

And what was it like working with NVGM? Very different than startup. Yeah, well in one hand the ASIC was very different. On the other hand, I was like an internal startup within GM because GM doesn’t have radar development department.

It was all built from scratch. I actually hired the entire team. I had the privilege of founding this and first engineering team in Israel.

And it was actually the first product out of Israel into cruise that was bought 2016. So I was lucky in this regard to put it in the vehicle. And so they bought cruise and then brought in your sensors following.

So GM bought cruise at I think March 16 for $900 million. And in my first demo of my imaging radar was September 16. So it was a crew CEO.

Kyle Vaught saw it and he said I want to go to my vehicles. December it was sent for integration. March it was on vehicles.

So I was lucky in this regard also to be in the forefront of the rush to Robotaxis back in 2016. We’re still rushing. We’re still there.

Yeah, well people don’t think it will happen. But back then the assumption was that in 2019 we’re going to launch tens of thousands of Robotaxis into the streets. I was ready with the production line for hundred thousands of radars a year for Robotaxis in 2018.

So I kind of in a matter of a few years I you know from starting this with two people into getting into a production line for Robotaxis. So it was real cool experience and I was lucky to be in the right time the right place with the right people to support me on this. So you’ve seen some really big moves and got to participate firsthand.

That’s very cool. Just for the record. I mean, definitely Gil Golan who was the Israel Tech Center manager, Eagle Billick who was the researcher here and in the States Andrew Ferrer and Scott Miller, you know great leaders and you know, it was a hell of a ride.

Yeah, that is awesome. Tell me automotive has changed a lot in the last 10 years and it’s still struggling to get adoption for electric adoption for autonomous. What is your take on the future of consumer automotive? Where are we headed right now? Yeah, well, I’ve always been in the front front of pushing this to reality.

I think one of my former leaders, he made a very bold statement which I support, you know, safety as a standard. I think all of us would like our vehicles A to be much safer B to do all those mundane tasks for us. But I think the automotive industry as a DNA.

I mean, it is a great DNA for a lot of things, but this requires a very dramatic reconfiguration of this DNA and we see all the audience trying either to do it in-house or with external bodies like carrier, etc. And but I don’t think the industry has figured out yet how to take those two, you know, very excellent doctors, right? One is a heart surgeon. The other one is a plastic surgeon, but somehow they need to work on the same patient.

And so I think it’s still I don’t think a lot of a lot of companies figured it out. I think companies like Tesla and Mobileye which are in DNA more of a deep tech software companies. It’s easier for them because you know to adopt the vehicle mindset is a bit easier.

But I think some will not survive some will be able to make this transformation, but it takes longer than expected. Yeah, I mentioned before that before we started recording I’m originally from Chattanooga, Tennessee and we had a big Volkswagen factory set there and when they broke ground on that I remember when that was about when Tesla was really starting to gain steam. We’ll just say 2015 ish plus or minus maybe even earlier and I was just like this guy is crazy.

What is he doing? Because Volkswagen they are German engineers very serious. The margins are incredibly thin. They take everything to the highest degree of perfection.

They’re looking at every paint coat looking for any blemishes like there’s no way somebody can actually disrupt this industry because it’s just incredibly expensive to get up and running and the margins are so so there’s just already thin margins and little did I know this wave of technology that was about to come out and truly the benefits of having these fully integrated systems and having a technology mindset. Now you see these other companies and it’s I can’t imagine trying to catch up from from a technology standpoint to somebody who is a tech first automotive. Yeah today.

It’s much easier. I agree with you completely and I think the one ingredient that was not there on the on the auto industry was software. Right and the minute that that became one of the key ingredients then and engines and vehicle design became more commoditized then then there’s a shift of balance between what’s critical for success and and it’s not so it’s not enough to bring a lot of software guys into your business.

You need a software mindset and this is not easy to adopt. Yeah, no, not at all. I think this is a thing.

This is a good moment to pause. I’m really grateful for you coming on the show. This is a another fantastic episode out of a cutting-edge company from a cutting-edge company out of Israel.

So we’ve had a few others and I’m just been really impressed and excited to see all the deep tech that’s coming out of there and really a pleasure meeting you. So thank you for coming on. Is there any way for anybody listening out there ways they can follow you or your company? Yeah.

Well, we have our site ArbyRobotics.com. Feel free to follow and we are pretty active. So well, thanks a lot again and we’ll connect again soon. Have a great day.

Thank you very much.

LEARN MORE: Discover the essentials of IoT connectivity with our comprehensive guide—an invaluable resource for anyone looking to navigate the complexities of the IoT landscape.