Podcast: Stanford & Toyota Drift Cars With AI

In this episode, we explore the groundbreaking achievement of AI-directed driverless drifting by Stanford Engineering and the Toyota Research Institute and discover how this collaboration is pushing the boundaries of autonomous vehicle technology, setting a new milestone in the field of AI-powered driving.

This podcast is sponsored by Mouser Electronics. 

Episode Notes

(3:27) - AI-directed Driverless Drift

This episode was brought to you by Mouser, our favorite place to get electronics parts for any project, whether it be a hobby at home or a prototype for work. Click HERE to learn more the current state of autonomous driving, the technical/regulatory limitations, and what the future holds!

Become a founding reader of our newsletter: http://read.thenextbyte.com/

Transcript

What's going on, party people? Welcome back to the Next Byte podcast. And today we're talking about drifting. That's right, drifting. One of Dan and I's favorite hobbies that we of course never do. But specifically, we're talking about how drifting assisted via AI can actually unlock a feature in autonomous driving that is safer for every driver out there, even those that don't like to drift. So, folks, strap in, buckle up, let's drift into it.

I'm Daniel, and I'm Farbod. And this is the NextByte Podcast. Every week, we explore interesting and impactful tech and engineering content from Wevolver.com and deliver it to you in bite sized episodes that are easy to understand, regardless of your background. 

Farbod: Friends, as you heard, today we're talking about drifting. I'm not gonna lie to you, this is a very, I guess, self-serving episode. I love drifting, Daniel loves drifting, we like cars, we wanna talk about cars. It's our podcast, and we're gonna do it. But…

Daniel: Legally, for the record, I don't personally love drifting.

Farbod: Oh yeah, yes.

Daniel: I love watching other people drift.

Farbod: Yes, of course, we would never.

Daniel: And playing video games with drifting.

Farbod: Of course, we would never go to an empty parking lot and try that out. Never, ever, not us.

Daniel: Not at George Mason University, in a way.

Farbod: Definitely not. Not in lot K, absolutely not. But before we get into today's episode, we're gonna talk about today's sponsor and that's Mouser Electronics. Folks, if you've been listening to this podcast, you know we love working with Mouser. And the reason we love working with them is because they relate really well with the things that we try to do. For example, telling an incredible audience that's interested in technology and autonomous vehicles and maybe even drifting about the cool things that are happening in that industry, in that realm. So, they write these incredible technical resources that are very easy to digest for someone like us that is interested and has some sort of a technical background, but also people that are just interested and don't have a technical background. They really know their audience, let's just say that. So, in the link that we're going to provide in the show notes for you today, they're talking about autonomous vehicles and the different levels there are out there. Specifically, they're referencing the Society of Automotive Engineers levels of autonomy. It goes from level one through level five. Now, if you've been in the US and you've been seeing most of these assisted driving technologies that's still required to put your hands on the steering wheel and whatnot, that's level two. Level three has been the next barrier to full autonomous vehicle driving. And there's some fuzzy area there that's between the legality of it and is the technology there. So, they dig a little bit into that, and then they also talk about the company's taken a stab at this. Specifically, they reference how in 2021 in Japan, they started doing these trials with Honda, and now finally in 2024 in the United States, Mercedes has taken a jab with the S-Class sedan and the EQS. So, if this field is interesting to you, if you wanna understand what Elon Musk actually means when he says next year we're gonna get full autonomous vehicles, as he says every year. Maybe this article can shine a light as to why it's been taking so long and what it's actually gonna take for us to get there. Yeah, that's the gist of it. With that out of the way, let's actually jump into today's article. We're gonna be going all the way to the West Coast of the United States to Stanford University. We got Professor Christian Gerdes, who's apparently just one of the biggest contributors to autonomous vehicles. Crazy, we've never covered him on the podcast, but now we got some homework to do. Out of Stanford University, working with the Toyota Research Institute. So, Stanford has, I think an entire, what is it? I'm blanking right now. Driving Circles course?

Daniel: Track.

Farbod: Track, there it is, oh my God. They have all track dedicated for just autonomous vehicle development, which is pretty cool. Up there with University of Michigan's little city that they have for just developing this technology. And obviously you got Toyota, which is one of the biggest and best automotive manufacturers in the world. And they've been taking on the challenge of making safe autonomous vehicles, right? It's no easy feat as we've been able to see over the past couple of years. And we already kind of spoiled it, but they're tackling drifting. And as a listener, you might be wondering why drifting? Dan, tell the good people, why drifting?

Daniel: Well, because there's nothing better in the world than drifting.

Farbod: Absolutely.

Daniel: But also, like the high-level thesis here is like if you can have AI that is reliable at the boundary cases in the most extreme driving conditions with precise high risk maneuvers that usually have to be performed by professional drivers, the thesis here is that you push AI to its limits in a controlled environment with drifting, you can advance the potential for safer autonomous driving in other complex scenarios that maybe aren't as likely as drifting, but again, you try and force AI and technology to be reliable at the boundary cases so that the not so boundary cases, you can take the lessons learned from that and make it more reliable.

Farbod: Exactly, and like the beauty of drifting, besides the aesthetics of course, is that you have to have this incredible combination and precision when it comes to the amount of throttle and the amount of braking that's being used. It takes a lot of finesse, you know? Like the driver has to have a lot of expertise executed the right way or else you're either not going to drift or you're going to spin out. Right?

Daniel: Well, and that's why it's not safe for folks at home to try this is it requires a lot of practice. You kind of have to like, it's the phenomenon behind it are so complex because the vehicle is slipping, but it also still retains a portion of its traction. The phenomenon is so complex that it's challenging for you to be able to like learn this sitting around in a classroom and then go immediately apply it on the road the way that you might with like turning on your windshield wipers or coming to a stop and then making a right-hand turn. Drifting is something that you almost need to entirely replicate by feel. So, it's a lot more challenging, right, to teach AI to do it. Cause you can't give it a bunch of strict rules in this they go do it. It also needs to kind of have a perception and a feel of, you know, and I'm saying feel and apostrophes here for folks who are listening instead of watching, you kind of need to help AI learn with feel instead of the fundamental phenomenon that underline, create the scenario of a car slipping, but still being able to control itself to some extent.

Farbod: Absolutely. And that sense of feel and just general vibes of the drifting happening the right way becomes tenfold when you're not just drifting by yourself but doing tandem drifting, which is where you have a leader and a chaser. So, you have one car that's leading the drift and another car that's mimicking it and following it in its path, right? Now you have some car, like let's say, hypothetically definitely not real, Daniel, you drifting in Laque at George Mason University, and then me drifting behind you and trying to follow your lead, right?

Daniel: This is actually hypothetical because there's no way we could actually pull this off.

Farbod: This is true, this is very true. But you know what I mean? Not only do I have to make sure that I'm executing the drift in just the right way so that I'm not spinning out or anything, but I also now have to keep your randomness in mind and follow your lead as you change your pattern. And like at times we could be what, a foot or two away from each other. So, the possibility of collision is actually quite high. So, you know, this is gonna require that right level of feel, which I feel like we've been teasing for a minute here. Let's jump into the sauce. Obviously, a difficult problem. We've tried our best to explain why it's difficult up to this point. Stanford University researchers, they're well aware of this and they kind of, through this whole idea that we gotta focus on a physics model, completely out the door. They're like, forget that. No, we're gonna go by vibes. So, we're gonna try to train a neural network that mimics and understands and grows over time as it drives so that it gets how to drift just the right way, right? And then, you know, I'm sure they use reinforcement learning to give it prizes as it does things well over time. What the outcome has been is that you have two cars, a chaser, I mean, a leader and a chaser, both being capable of drifting at up to 35 miles per hour speeds, at times 10 inches away from each other. So, like just at the level of finesse that you would expect from professional drifters, these two cars are able to execute together.

Daniel: It's crazy. And I mean, it's not quite the same. This isn't tandem drifting, but it reminds me of the scene in Fast and Furious Tokyo Drift. Do you know when I say the scene?

Farbod: I know the scene.

Daniel: It's the one where there's like, there's two girls sitting in the car and then the other guy, the two guys come at them in the Mazda and they're driving straight at them. And then they like drift a circle around them and it's super close. And the margin for error is so tight that if either one of those cars were to move unexpectedly, the entire thing could have turned into a crash. This feels like that, you know, a mix of that Tokyo drift, fast and furious-esque, mixed with like watching Formula One Grand Prix racing, where these vehicles are like inches apart from one another, traveling at super high speeds, snaking their way through a course. And you know, the chase vehicle has to be able to react to changes in courses for the lead vehicle without causing a crash. And again, just as a, as I'm, I don't know if you can tell, as a slight motorhead, as a car enthusiast, it's something super beautiful and awesome to watch. We've got, we'll have this article linked in the show notes. I feel like this is like mandatory watching for anyone who's listening to this podcast.

Farbod: So beautifully shot. So incredibly beautiful shot.

Daniel: Go check out this article linked in the show notes. Watch it's three and a half minutes long. And about one and a half minutes of that is just cars drifting around each other. So, it's definitely worth a watch.

Farbod: Definitely. And, another thing that I realized now I didn't mention, what was beautiful to me is that in terms of the underlying technology of, how the leader is path planning and how the communication is happening in general, they're just using GPS and Wi-Fi, which is not foreign. Like most of these modern new cars coming out. And I'm going to reference Tesla. They already have that technology on board. This is already in place. So, a lot of the juicy bits happening here are with the neural network that's operating the vehicles. So, you have the one upfront doing the path planning, the one behind avoiding the collision and following its lead. And obviously what this has culminated to is a great drifting experience. But for the average listener, why does this matter? You know, like when I'm driving down Route 7, I don't care about drifting. And that's most people. Well, what you do care about is whenever we do reach level three, level four, level five autonomy, you wanna make sure that when you take your hands off of those steering wheels, your car will not collide with another car, especially in these unperfect conditions. We know what Maryland drivers and Virginia drivers are like. To the folks outside of the United States or out of the East Coast, they're very bad drivers. That's the running joke. They're horrible drivers. So, I guess make up for those uncertain scenarios where someone is doing something bad, having a model that is so good at behaving in an unpredictable scenario that it can drift is gonna come in quite handy. And that's the main motivation behind what they're doing here, not the beautiful, amazing drifts that you can watch a minute and a half of on the length video.

Daniel: Well, and I think, again, like the high level here is drifting is a perfect analogy for normal human being driving on ice, right? Lack of friction is the phenomenon that we're trying to understand here. Drifting is self-induced, purposeful lack of friction, but you know, you or me driving on route seven on our way to one of each other's houses, if there's an icy patch in the road, the same fundamental principles for drifting apply again here. And again, that's where we talked about creating deliberate boundary cases to test and to train AI in a way that makes it more bulletproof, more reliable for the common person because those boundary cases are boundaries for a reason. They don't always occur, but when they do, you want a way of having generated these cases in the past to train the AI algorithms to be really, really robust. And I think one of the cool things to talk about here is just how fast and how heavily developed this AI model and how many calculations they do. They apparently are able to re-plan their route up to 50 times per second. They're instrumented with a bunch of sensors that essentially allow the vehicle to feel gravity and to feel friction almost as a metaphor for the way that a driver would feel it, if they were inside the car driving. So, I like the idea here that this entire power train, transmission suspension are fitted with an array of sensors and computers that control the steering, control the throttle, control the brakes, obviously, right? It's self-driven, but then also have control mechanisms that help it gain feedback on how is the vehicle moving? What does the inertia feel like inside the vehicle? Because that's how a human driving the vehicle would feel and use their senses to try and control it.

Farbod: Absolutely. Absolutely. Now, one thing that wasn't mentioned in the article and I wasn't is how much of this model that they've been fine tuning and developing is constrained to the vehicle that it was trained on. Basically, like how much of it is portable and can be applied to a different car. I remember I think a couple weeks ago we did an episode on a robotic platform where it had like general knowledge of how to do a task or a set of tasks and then the last 20% of the learning would happen you know on the job where we'd get it right. So, I wonder if it's something like that, where it's just the general idea of what it means to drift the right way is like same across the board. And then the feel of the car that has to be factored into those measurements is then the last 10% that might have to be fine-tuned.

Daniel: Well, I'm with you in it. I would gladly volunteer to test this in the vehicles that they developed it on. I feel like it'd be fun if you and I were doing tandem drifting together in the pair of Supras. They took Toyota Supras and that's the ones that they did the development on. But I'm with you, it would be interesting to understand not only specific to the vehicle model of the Supra, but also all the instrumentation available on the Supra. Is this something that you can't trust a normal self-driving car in the road because it doesn't have enough sensors? Or are the sensors that they fitted this Supra actually pretty equivalent to something like what a Tesla or a Waymo taxi might have. Like I would be interested in understanding are the capabilities of this car vastly different than your run of the mill autonomous vehicle.

Farbod: Yeah, that's a good point. That's a good point. Yeah, I think that's really all I had on this.

Daniel: I do have something else to mention.

Farbod: Sure.

Daniel: Just because I have to give some kudos to Top Gear. So, I was doing some research for this episode. Top Gear the blog, not Top Gear the show, which is by the way, one of the greatest shows of all time.

Farbod: Amen to that.

Daniel: The Top Gear blog, I was doing some research on this, read one of their articles related to this development. The headline, Toyota completed the world's first autonomous twin drift in a pair of Supras, but the subheader here is like the most important part. And I think it completely captures our feelings and emotions on this, right? So, this technology yields positive impacts on the development of future terrain talk technology cool. But mostly look a self-driving twin drift like that's where our heads are at like this is super cool. It does have potential implications for like driving on ice driving a mud driving in low mu or like low friction conditions. But also look it's really cool. These two cars are independently autonomously controlled and they look like a pair of figure skaters dancing around each other. But instead of it being on ice, they're on normal asphalt. I don't know, it's awesome.

Farbod: It brings me back to the first time I watched Tokyo Drift as a kid. It's just, there's something magical about it. I can't explain it. It's incredible.

Daniel: Yeah, it evokes some level of nostalgia. I'm with you, man. And I will, like, I'm just gonna take a little bit of a jab here. I feel like we get a lot of like, I see hype videos on Twitter about like, wow, like look at my Tesla autopilot. Like it's switching lanes for me on the highway, which is like a really routine thing that everyone does every single day. And I understand that that is valuable and that is useful, but I feel like companies like Tesla and Waymo focusing on like basic road safety. This project is focusing on the complete opposite end of the spectrum, pushing the limits, making sure that if self-driving cars are safe in a twin drifting scenario, which is probably the most dangerous driving you can possibly do. I'm trying to take the learnings from that and apply it to like, you know, your run to the mill merging on a highway.

Farbod: Shots were taken. You just, my God.

Daniel: That's not to save, you know, a Tesla or a Waymo can't make their car drift. I just, I appreciate Toyota's research Institute kind of taking the fun way on this and their Japanese OEM, Tokyo Drift.

Farbod: Supras. You know, like.

Daniel: There's a little bit of favoritism here if you guys can't tell.

Farbod: Yeah, it's coming full circle. All right, let's do a quick wrap up. Folks, if you ever watch Need for Speed, Tokyo Drift, and thought, I hope one day I could do that, and you know, likely you never did and you never will, well, you're in luck because researchers at Stanford University have come up with a neural network, an AI algorithm that can drift. And not only can it drift. It can tandem drift. That means you and your best buddy can drift together in parking lots where you probably should not be drifting. Now, the reason this is important is because they've been working with the Toyota Research Institute to figure out how we can make AI models, remember when we're getting to level three, level four, level five autonomous driving, that we can really trust in uncertain scenarios when driving in the road. Imagine if it's really icy. That's kind of like how you'd be managing your car if it was drifting. It's really cool what they've done here. They've kind of ditched the physics model and went completely on the feel of the road and gravity that a driver would usually feel. And the result is astounding. They've been able to accomplish drifting at 35 miles an hour, two cars next to each other, sometimes 10 inches away from each other without crashing, drifting like absolute pros. That's why I think this is the biggest news in the automotive industry in the past god knows how long.

Daniel: Nailed it.

Farbod: I try. I do what I can, you know, for the friends, for the drifting, loving friends.

Daniel: You know, I'm hoping one day, Stanford friends, if you're listening, Toyota Research Institute friends, if you're listening, let us know, reach out. We would absolutely love to hop on a plane. Just let us know. Well, we would love to hop on a plane, film some freaking sweet content with you guys, post it all over our social media, cause I'm sure our community would be just as excited about this as we are, at least a portion of them.

Farbod: In fact, let us know. Comment, let them know, tag them, let them know that we wanna do this just for you, you know? We wanna make content for you, it's not for us.

Daniel: Yeah, if you wanna see us ride in autonomously drifted cars.

Farbod: Give us a shout.

Daniel: You wanna see us die.

Farbod: Give us a shout. All right folks, thank you so much for listening and as always, we'll catch you in the next one.

Daniel: Peace.

As always, you can find these and other interesting & impactful engineering articles on Wevolver.com.

To learn more about this show, please visit our shows page. By following the page, you will get automatic updates by email when a new show is published. Be sure to give us a follow and review on Apple podcasts, Spotify, and most of your favorite podcast platforms!

--

The Next Byte: We're two engineers on a mission to simplify complex science & technology, making it easy to understand. In each episode of our show, we dive into world-changing tech (such as AI, robotics, 3D printing, IoT, & much more), all while keeping it entertaining & engaging along the way.