Matt Kixmoeller
Welcome to Talking Autonomy. This episode's all about radar, or I should say software defined radar. I'd like to introduce you to Tegan, who's joining us today and leads our Radar Center of Excellence in Dallas, Texas. Tegan, tell us a little bit about yourself and what you and the folks down in Dallas do for Ghost.
Tegan Counts
I'm a radar systems engineer and with an Electrical Engineering degree, and I spent the last 15 years before Ghost in the defense industry, mostly making missile seekers radars. Ghost opened the Radar Center of Excellence in Dallas because of the strong talent base for radar engineers in that area, again, from the defense background. We brought together some engineers from defense, some engineers who had experience in the autonomy space, and we stood up a radar team whose goal is to make the next generation radar to support Ghost.
Matt Kixmoeller
It must be super confusing for 15 years of your life to be working on radar and trying to seek things out and now you're trying avoid them! But other than that, give us a sense for how the technologies of military radar translate over to the auto world.
Tegan Counts
There's a couple important technologies. The first one is interference mitigation. In a defense radar, you're expecting that your adversary is going to send out some jamming signal and you're going to be in a very cluttered electromagnetic environment. In autonomy, although nobody's malicious, the environment is still very cluttered. More and more cars are equipped with radars, and those radars all provide signals that are interfering with the one that we're trying to send out and receive. That interference can lead to false alarms or missed detections. The second thing we bring from defense is multifunction radar. Multifunction radar is constantly changing the mode that it's operating in based on what it senses in the environment and what the autonomy system needs it to do. If we didn't have a multi-mode system, we would've either had to build hardware that was excellent across all dimensions and that would be too expensive, or you would do what people do today and that's build hardware that is modest across all dimensions. But with multifunction, we can take that same hardware and employ it differently over time. And when I say over time, I can mean even over millisecond intervals. We can be changing what we're doing in the hundreds of milliseconds. We can go from a high dynamic range mode or a long range mode, we can do different things with our PRFs. Going back to interference mitigation, we can be constantly sensing and adjusting to the environment.
Matt Kixmoeller
One of the key elements of our design is that the radar is software defined. Break that down for us - what does it mean to be a “software defined radar?”
Tegan Counts
When you're a software defined radar, you are actually passing back the raw signal that comes off of the A-to-D converters into a software stack that you can change over time. In a traditional setup, you would have your signal processing living right up next to those A-to-D converters, you may have an FPGA or a custom ASIC sitting right behind it that would take in those signals and process them down to a much smaller volume. Then you'd put that on a low bandwidth data pipe and send results to the central computer. What we're talking about is actually sending that higher bandwidth data from the radar front end back to the central computer for signal processing. We can do that with commercially-available technology so it does not drive our price up, and it gives us this tremendous flexibility of having all of our algorithms in centralized compute. First of all, that's a cheaper environment to have a processing resource live in, so it helps control cost. It also helps with flexibility because now I can be constantly evolving my algorithms in real time, shoulder to shoulder next to the system engineers who are working one level above us. So that's where that interaction and ability to dynamically change your data products to suit the system need comes into play.
Matt Kixmoeller
That sounds amazing, and I'm struck by the parallels with our vision system. We figured out early on that if we could just have raw camera feeds sending data back to the central processor and enabling neural networks to see all the information, we could do much more with it. We're essentially now taking the same approach with radar where we're trying to get as much information as possible centrally so we can analyze it to the full abilities.
Tegan Counts
That's absolutely right.
Matt Kixmoeller
I feel like in general, it's a bit of a renaissance time right now for radar. A lot of people are excited about it and I think that stems from a belief that we can elevate radar to be a primary sensor. Many of the earlier deployments of radar in autos used it with a camera and you really needed have a camera to find an object to be able to cut through the noise of the radar to isolate objects and find their distance. But with some of these new techniques, we can really elevate radar to a primary sensor that's a peer with a camera, do I have that about right?
Tegan Counts
It's understood that if you use basic algorithms and basic detection techniques, you're going to have too many false alarms. This is expected, if you will, which is why camera confirmation is important but it's not actually a limitation of radar. Radar itself is capable of providing a clean enough signal, but it requires a different hardware and most importantly, different software. Those techniques are all born and proven in the defense world. We'll bring those techniques to the auto industry, and now we'll get a clean signal that the autonomy system can use and trust that is not laden with false alarms or mis-detections.
Matt Kixmoeller
One of the topics you brought up in your intro was about multi-mode and how we're developing the radar in a way that we can change in real time how it's behaving and change where it focuses. Say more about that please.
Tegan Counts
When we do this, we have the ability to decide which dimension we're going to maximize at that point in time, dimensions like angle, accuracy, range, dynamic range, interference mitigation, et cetera. So by interacting with the system and by us sensing and probing on the environment, we can choose to maximize differently at different points in time. It's going to be constantly switching between modes as it goes, and that's going to allow us to give the performance in the wide range of environments without having to expand the cost. That's why it's a big deal for us.
Matt Kixmoeller
So if I'm driving down the highway and we're going, you know, 60 miles an hour versus in stop and go traffic, we might change the mode.
Tegan Counts
I would expect you to be in definitely a different mode between those two situations, for sure. We'll see the density and the traffic will change. How will we know to change modes? It's the behavior of the car, the behavior of the environment around us, and we'll use those two factors to decide what mode we need to be in.
Matt Kixmoeller
Wow, I'm really getting a sense for the intelligence that software defined can bring it. It also allows us to configure the radar differently for a forward application versus say a rear or a side application as well, right?
Tegan Counts
Absolutely, you have different physics and different kinematics, so you'll tune them in different ways and even forward and side, they'll each be doing their own thing separately that's optimized for the mission they're trying to accomplish at that moment. So that's why each of the different radars can operate in different modes and we'll make sure we don't have any mutual interference between all the ones that are on our cars.
Matt Kixmoeller
People are often surprised that we dove in and decided to build our own radar. Why take on that ambitious project versus just taking one of the off-the-shelf radars that are on the market today?
Tegan Counts
We actually evaluated a lot of off-the-shelf radars to see if we could make them work, but they all had the same limitations. They produced data products that were not based on what the system needs. There's a radar company building a radar and they hope that one day an autonomy company will pick it up and do something with it, but it doesn't allow that mutual interaction. It doesn't allow that evolving of the requirements between the system and the subsystem. The other limitation that we found when we were looking at these off the shelf radars is they had their ASIC-based onboard signal processing, and it was very difficult for us to get in there and change it. We knew fundamentally we had to go change the way the signal processing worked.
Matt Kixmoeller
We wanted to be able to tune it and really customize it for our needs.
Tegan Counts
That's right, and that was just not going to be possible with those off-the-shelf radars.
Matt Kixmoeller
The design of the hardware is, as I understand it, really nothing special, all the magic is in the software. How were you able to achieve these interesting outcomes using relatively off the shelf hardware?
Tegan Counts
We'll often hear people say ‘well you can't get level four with this, this is a level two part.’ Well what does that really mean? It's more about how you use the parts. What we have are the core are TX/RX chips that are widely available in the market. What we've done is put a different antenna topology in front of them, we have different calibration techniques, we have different algorithmic approaches, and we bring all those things together - that's what actually unlocks the performance. There is not a need to go to a very expensive or a custom ASIC solution to the front end because those aren't actually going to solve the problems that you have, especially in the domain of interference. You can't just throw hardware or money at the problem. You actually have to have the smarts in your software, in your algorithms.
Matt Kixmoeller
Add it all up for me, I mean, what are the dimensions by which this performs differently or better now?
Tegan Counts
You're talking about better angle accuracy, you're talking about better dynamic range, especially in the angle dimension. And then again, most importantly you're talking about signal interference rejection ratios which are tens of DBs better than what you're seeing in these typical radars.
Matt Kixmoeller
Another big area that you focus on is calibration. Why is calibration so important for radar and what have you done there?
Tegan Counts
What calibration allows you to do is sense features in the environment more finely. In a typical system, maybe you can only see things that are within five dB of one another, or if you want to talk about spatially, maybe they're within just, you know, a few tenths of a degree of each other. With better calibration, now we can separate those things to an even finer and finer degree. Again, this comes into play for interference mitigation. We can sense something in the environment and we can very quickly know if that is interference from another vehicle or if that's a true return from another vehicle. Calibration is what leads to performance, but calibration can also be very cost prohibitive. If each radar has to spend hours or days in a calibration environment before you put it on a car, and it very delicately falls away from that calibration, then it's not going to be cost effective. What we've done from the beginning is architected our calibration techniques so that they will be fast to run in a production environment, they'll be stable once they go on the car and we'll be constantly doing online calibration assessments and adjustments using the radar and the camera and the rest of the system, so that our system will stay stable versus time.
Matt Kixmoeller
That's really cool, I mean, not only can we update the software and improve the capabilities over time, but you have this ongoing calibration that makes sure the device can last in a car for a decade-long deployment or more.
Tegan Counts
That's absolutely what it has to have. All the sensors talking together, feeding that information back, we'll actually see as the cars are deployed, we'll get that information back here at Ghost, and we can look for then patterns. Patterns across cars, patterns across lots, across radars. All those things, all that data that we have access to is how you make a smarter radar system.
Matt Kixmoeller
Tegan, that sounds amazing. Thanks for leading this ambitious program for us, a new approach to a software defined radar where we take all the intelligence, bring it centrally, and allow our engineers to build functions that take advantage of radar and vision together to have the best capabilities to power our driving pipeline.
Tegan Counts
Thank you.
