Skip to content
Driving Car

Can Driverless Cars be Trained Virtually?

🎙️ Enjoy our PODCAST on this topic

* podcast transcript

EVA
Self driving cars, right? They’re kind of that thing, you know, everyone dreams about like straight out of sci-fi. But then let’s be real, they also make people kind of nervous and think about it. Being able to just zone out on your commute, answer some emails or just like finally enjoy the view, you know, sounds amazing, but then there’s this other party going. Hold on, are these things really ready to be out here with the rest of us? And that’s kind of the big question, isn’t it? Can we actually trust algorithms with, well, our lives?

MAX
And can we get to that driverless future, you know, without waiting?

EVA
Exactly so we’re diving into this article. It’s called can driverless cars be trained virtually to, you know, see what’s up, see if the digital world holds the key. Now here’s the thing. One of the biggest things holding these self driving cars back isn’t some crazy tech stuff. It’s just plain old mileage. The Rand corporation, they did a study and it turns out to be really sure these cars are as safe as us. They’d need to drive a ridiculous amount, like 11 billion miles of data.

MAX
11 billion? That’s like what, driving around the earth like over 400,000 times?

EVA
You got it. Even if you had a ton of these cars driving nonstop day and night, it to take like centuries. So unless someone’s got a time machine line around, real world testing alone just ain’t going to cut it.

MAX
And you know, what’s so interesting about this whole 11 billion mile thing is it forces us to like, really rethink how we’re approaching this. It’s not just about, you know, racking up miles. It’s about what those miles represent. Every snowstorm, someone walking across the street when they shouldn’t, the sun reflecting off something and messing with the sensors, those weird unexpected things, those are what really test these AIS.

EVA
So how do you teach a car to expect the unexpected, right? That seems like a pretty crucial skill for, you know, driving. Well, that’s where things get really interesting. And this company, Parallel Domain comes in. They’re basically building this like Ultimate Driving Simulator. We’re not talking about those old janky driving games. We’re talking crazy, realistic virtual worlds where every little detail matters. I’m talking like city blocks that can create those in minutes, not months. Remember those old driving games where if you went too fast you’d outrun the graphics and like hit the edge of the world? Parallel domain is on another level.

MAX
Yeah, and this is where it gets, you know, super interesting. It’s not just about speed. By making virtual testing more, I guess accessible is the word parallel domain is kind of shaking things up. Smaller companies, they could never afford to put a bunch of self driving cars on the road, right? But now they can compete with the big guys, the Teslas of the world.

EVA
OK. So I see where you’re going with this and it’s kind of huge. We’re not just talking about making these cars, we’re talking about like changing how cities are designed. Imagine cities that are built for these self driving cars. You know, where the cars can talk to each other and everything.

MAX
Exactly. And and think about the safety aspect. Oh right, testing in these virtual worlds means no one gets hurt while we’re trying to figure this whole self driving thing out. Let the AI make its mistakes in the virtual world, not out here with real people.

EVA
That’s a really good point, but let’s be real. Virtual reality is one thing, but getting people to trust a car that learn to drive by playing like a really intense video game, that’s another thing entirely.

MAX
It all comes down to who’s behind the code, right? And that’s where Parallel Domain’s founder, Kevin McNamara. He’s got this really interesting background. We’re not talking about some random tech guy. Oh, this is a guy who made his name creating virtual worlds. But get this for Pixar, hey?

EVA
Pixar like Buzz Lightyear, Toy Story, that Pixar that’s…

MAX
… the one and Microsoft games.

EVA
No way. The same Microsoft that brought us what, Halo? OK, now that’s some serious St. credit.

MAX
Exactly. So when this guy talks about simulating reality, he knows what he’s talking about. It’s not just about, you know, making things look pretty. It’s about those details, the light, the way things move, the little things our brains pick up on that make us believe it.

EVA
So they’re using the same tech that makes us believe a toy can talk to teach a car to drive. That’s that’s next level. But how does it like actually work?

MAX
It’s all about data. Real world maps, data from sensors, they take all of that and feed it to these things. They’re called generative algorithms.

EVA
Generative algorithms. OK, so for those of us who don’t speak fluent tech, what does that even mean?

MAX
Good point. They’re like, imagine a really, really advanced recipe book, but for computers. You give it the rules, things like the laws of physics, how pedestrians usually act, you name it, and it takes those rules and comes up with like endless possibilities.

EVA
Like it’s simulating every possible traffic jam, every unexpected detour. So they’re basically stress testing these cars in the Matrix before we ever get in. That’s that’s both amazing and slightly terrifying at the same time, but it still sounds kind of theoretical. Are there actual companies using this? Like for real?

MAX
Oh yeah, for sure. Actually their first customer parallel domain. It’s this Chinese electric car company NIO.

EVA
NIO.

MAX
And and they want to get into the US market.

EVA
Wow, going from like virtual test drives to American roads. Talk about a trial by fire. That’s that’s kind of bold, right? What’s the strategy there? Are they just trying to get attention or is there something more to it?

MAX
I think it’s a risk, definitely, but it could really pay off. They’re not just, you know, dipping their toes in the water here. They’re going all in. They’re betting that this virtual testing thing, it’s going to give them an edge, prove their tech is ready, and if they can make it in the US, well, that changes everything for everyone.

EVA
So NIO, they’re like the first domino, if they make it, the whole industry might shift towards this virtual testing.

MAX
Interesting. And what about parallel domain? What’s their, like, ultimate goal here? Their founder, McNamara, he says it’s like they’re building a Fast Forward button for the whole self driving thing.

EVA
The Fast Forward button.

MAX
Yeah, they don’t just want virtual testing to be like a common thing. They want to make those millions of miles. Remember we talked about those. Yeah, they want to make those almost not matter.

EVA
So instead of waiting years, decades even, they’re saying let’s do it all in the virtual world way faster. It’s a pretty big vision. It is, but I mean, there’s got to be more to it than just like hitting play on a simulation, right?

MAX
You’re right, it’s not quite that simple. And that’s where even the best simulations, they kind of hit a wall. You could try to recreate everything in a computer, but there’s always that, you know, random element, that thing you just can’t predict in the real world.

EVA
Right. Which brings us to that Uber accident. It’s been a while now, but it’s still like this big thing hanging over everything. It reminds you that even with all this amazing tech, there are still real risks.

MAX
Absolutely. That’s why virtual testing, it’s just one part of the whole picture, not the whole thing. It’s about balance, you know, use the simulations to learn faster, deal with the stuff you can predict. But real world testing, that’s always going to be necessary.

EVA
Because you can’t simulate everything.

MAX
Exactly. Especially not how people act.

EVA
So we’re moving towards a world where driving experience, it means something totally different. It’s not just about miles on the road anymore. It’s about those virtual miles too.

MAX
Exactly. And that brings up a really interesting question

EVA
What’s that?

MAX
How do we even measure those virtual experiences? How do we know what they’re worth?

EVA
Right, like if a self-driving car can handle a million virtual rush hours without a scratch, does that make it a better driver than one that’s been out in the real world a bit but not as much? How do you even compare those two things?

MAX
It’s tough. It’s like, it’s like comparing, I don’t know, maybe an apple you’ve only seen in a picture to one you’ve actually held, taking a bite out of both apples. But the experience, it’s just not the same, you know?

EVA
And that’s where things get kind of, I don’t know, trippy, right? Like if these simulations keep getting better, more real, what does it even mean to be a driver?

MAX
It makes you think, doesn’t it? And it’s not just about cars, either. Think about flight simulators. Pilots have been using those for ages to practice like crazy situations. But at a certain point, does that virtual experience become as good as actually flying?

EVA
It’s like that. What’s it called? The thought experiment If you could just download driving experience into your brain Matrix style, would you be a better driver than someone who actually spent years on the road?

MAX
Right. And how would you even like measure that? What makes someone a good driver? The reactions, the decisions they make, being able to predict what other people are going to do.

EVA
It’s a whole different world, that’s for sure. And it’s not just driving, it’s everything. Think about how much we rely on algorithms already, the news we see, who we’re friends with online, even the way we get to work.

MAX
That’s almost like you’re all living in our own little simulations, you know?

EVA
Whoa.

MAX
And the lines are getting blurrier all the time.

EVA
So here’s something to think about as you head back out into the, well, increasingly simulated world. If we’re handing over more and more control to algorithms, letting them decide everything, what does a real experience even mean anymore? Is it just like what we feel, or is there more to it, something even the best AI could never copy? That’s all the time we have for this deep dive, but the conversation definitely doesn’t end here. Until next time, stay curious everybody.

Self-driving vehicles are the future. From Tesla’s fully electric self-driving semi to Intel’s Mobileye assisted driving system, vehicle manufacturers are embracing this rapidly advancing technology.

But is the technology really perfected enough for us to see fully automated cars on roads today?

The RAND Corporation recently asked, “How many miles of self driving would it take to demonstrate autonomous vehicle reliability?”

They found that self-driving cars need to travel at least 11 billion miles of test time before complete confidence in safety can be reached. Given that large number, that can translate into tens or even hundreds of years of testing.

Analysts and researchers agree some innovative testing strategies are necessary. Typical, real world testing just isn’t practical to reach the time needed to say definitively that self-driving cars are safer than those operated by humans.

Finding Alternative Testing Methods

Companies are now finding ways to model environments to test assisted driving systems before they are testing in actual cars.

Notably, Parallel Domain is working toward accelerating training times by creating software that builds simulations quickly. Testing is difficult to coordinate while keeping safety a top priority. Creating a virtual testing environment makes pedestrian and passenger safety a non-issue.

Vehicle companies creating their own simulations can be cost prohibitive and extremely slow. Developers can take weeks to create a few city blocks. However, Parallel Domain was able to create a simulation that creates exquisitely detailed, hyper-realistic city blocks in minutes.

This is the solution the autonomous driving industry was looking for.

Waymo self-driving car side view.gk
By Grendelkhan [CC BY-SA 4.0], from Wikimedia Commons

Parallel Domain isn’t the only company bringing their A game to the self-driving car industry. Alphabet’s Waymo division has cooked up Carcraft, a virtual world were autonomous cars learn to drive. In 2016, their team logged over 2.5 billion virtual miles on miles of complex maps modeled after Austin, Phoenix and Mountain View.

Unlike Parallel Domain, Waymo made this proprietary software to serve their own company’s fleet of self-driving cars. Therefore, unlike Parallel Domain, Carcraft is not available for use by other car companies. It seems Parallel Domain’s founder may have invented a niche service after all.

From Pixar to Large Scale Modeling for Autonomous Vehicles

Parallel Domain was founded by Kevin McNamara. He began building virtual environments for Pixar films and later at Microsoft Games for video games like Crackdown 3 and Sunset Overdrive. After his time there, he moved to Apple’s Special Projects Group working on simulations for autonomous software.

When McNamara saw an opportunity, he went for it. Through Costanoa Ventures and Ubiquity Ventures, he managed to raise $2.5 million to start his new company. According to McNamara in an interview with Tech Crunch, “What we do is use computer graphics to try to accelerate the development of safe autonomous vehicles.”

The point is to do this all virtually in order to safely make mistakes and learn from them.

Essentially, Parallel Domain’s software uses real-world data from maps, along with generative models and growth algorithms, to create a virtual environment. Here, autonomous systems can learn to drive. Every aspect of the software is programmable and can be changed by researchers. Everything from terrain to road curvature is customizable. Pedestrians and time of day can be adjusted with the click of a mouse.

Parallel Domain already has their first customer: NIO, an auto driving vehicle startup out of China. The company currently only markets their vehicles in China. By 2020, however, they plan sell their cars in the United States. The cars are expected to only have some autonomous capabilities (think Tesla). Soon, other companies will likely join NIO in using this first of its kind technology.

The long-term vision by McNamara and his team is for Parallel Domain to work in tandem with real world tests. The simulation would act as a fast-forward button for testing. Car companies would be able to order a “package” simulation that includes millions of miles of diverse roads to drive on.

For auto companies, they would have the benefit of accessing large testing environments at a fraction of the cost. Thereby, more autonomous vehicle companies would have a chance against large competitors like Tesla to get their ideas tested. Software like Parallel Domain enables companies to develop products safely and quickly.

“Our software automatically generates the environments and scenarios that feed into simulators, making it safe and fast for autonomous vehicles to learn from their mistakes, accelerating time to safety for all vehicles.”

Kevin McNamara, CEO at Parallel Domain

With software like Parallel Domain’s, reaching 11 billion miles will be an attainable goal.

A Final Question

One thing needs to be considered: Will the public really trust virtual training in self-driving cars? In March of 2018, one of Uber’s autonomous cars was involved in a fatal accident – the first in history. It damaged the credibility of the industry as a whole.

Although creating simulations to log and test autonomous driving systems is essential, it’s no substitute for real life driving. It’s clear that a multi-level approach to training will be both the most accepted and most efficient way to get self-driving cars on the roads faster.

Self-driving car companies and simulation makers seem to understand the importance of virtual and real-world testing. Let’s hope it goes a long way in creating trust in consumers.