Self-driving cars are still going to take a long time, people

A few years ago, I visited a super promising company called Luminar, which makes LiDAR, or Light Detection and Ranging, technology that helps autonomous cars see the world around them. Now, Luminar is going public. It plans to enable autonomous driving on highways by 2022.

It’s kind of like what Tesla promises with Autopilot, except less human intervention required. I spoke with Austin Russell, CEO and founder of Luminar. He told me getting self-driving cars working in cities is hard, but autonomous highway driving is easier and cheaper. The following is an edited transcript of our conversation.

Austin Russell: You’ll actually be able to get on a highway with this, enable autonomous mode, take your hands off, eyes off, read a book, use your phone, work on your laptop, watch a movie, take a nap. And then a couple minutes before the final exit, have a planned manual takeover to drive to the final destination. We’re talking costs on the order of $1,000, as opposed to $100,000 for these kinds of autonomous vehicles. That’s something they can actually put into a consumer vehicle, something that you could buy.

Molly Wood: This approach, though, has also been referred to as kind of the mushy middle of automation. We’ve already sort of seen people over-rely, let’s say, on technology, like Autopilot — not take control when they need to. This feels like it almost even increases the opportunity for that. If I straight up go to sleep on the freeway, and then the expectation is I’ll wake up in time to take my exit, I can kind of see where that could go wrong in a hurry.

Russell: I mean, the main difference between a level two assisted driving system and an automated driving system, like in this case, it has end-to-end functionality within a given operational domain. That’s really what is there. Ultimately, it’s up to the OEM (Original Equipment Manufacturer) to determine the specific functionality and domain that is specifically determined and fully tested and validated over these various use cases to be safe in. It doesn’t necessarily mean that initially every car that you buy will be able to convert, like a transformer, into a sleep pod.

Wood: I understand the technical difference that you’re describing between autonomy and assisted driving. I suspect most consumers will not, and we actually know that most consumers use assisted driving in a way that is, frankly, closer to [how] drivers would behave if they were expecting full autonomy. But if it feels like a distinction without a difference, and then anything that goes wrong is really highly publicized, does that ultimately hurt adoption? Does taking a stair step approach make the long-term adoption less likely because there’s a lack of trust?

Russell: And the answer is, as it relates to systems today, right now, absolutely. I mean, the amount of folks that are calling assisted driving systems autonomous right now, I’d say — it actually happens more often than not, described inaccurately than accurately. And it will, and has posed challenges, for sure. I mean, even for the various accidents that have happened today, all of these systems are because they rely on a backup driver to be constantly paying attention, because they would be getting an accident left and right, if there wasn’t anyone paying attention. They don’t pay attention for some finite amount of time. That happens to be that time when they did need to be. And then it gets in an accident, and either leads to some kind of serious injury or fatality, which is extremely unfortunate. It gives the whole autonomous vehicle space a bad name for those things. It’s like, “Okay, well, why didn’t it prevent it?” Well, it wasn’t designed to do that. That was not the purpose of what that technology is in this current state. So there needs to be a very clear and distinct difference.

I mean, that’s why when we’re talking about the deployment of this kind of system, it really is, and has to be before it’s deployed — otherwise, it just won’t be deployed — safer than human level capability in the operating domain that is defined to be. And there has to be very clear lines, very clear distinction. I actually don’t think, to be frank, there’s a huge use case for the level two assisted driving systems that are out there today. I think it’s, to be frank, a little bit gimmicky in the sense that it kind of, as you said, lulls you into a false sense of security that you can kind of take your hands off, relax. The reality is, just like when you’re driving at any other point, anything can happen at any moment, [and] you have to be ready to take over at an instantaneous notice. On average, it’ll maybe only give you a second heads up before you actually need to take over. So I think that’s where there needs to be a fundamental shift. There needs to be a very clear distinction. And when you’re operating in autonomous mode, with our sensing system and software on the car, that’s where it on average would be safer than human level capability. That’s the baseline level of what’s needed to be able to see this through on these vehicles in that respective domain.

Wood: I feel like, when we talked a couple of years ago, you were like the guy who was going to fix it all. And I wonder, why not just kind of go all the way? Why not be the person who solves the urban autonomy problem? Why not release the full end-to-end solution?

Russell: The way that I see it is that this is the entryway into the long-term vision around level four or five robotaxis. The main challenge, though, is there has to be a commercially viable application of the technology in the near term to be able to really ensure this industry gets successfully realized. From a database perspective, if you want to be able to successfully deploy something outside of one specific city, you’re going to need a huge amount of high quality data, including this 3D LiDAR data everywhere around the world. We’ll have our own fleet out on the road, giving data back, to be able to better inform the system, and that’s the kind of foundation that you can build a global scale urban robotaxi system off of, not just in the near term, but the long term.

Related links: More insight from Molly Wood

MIT released a study last week that found that Tesla drivers definitely are not paying enough attention when they’re using Tesla’s Autopilot mode, which is supposed to be assisted — not autonomous — driving, as in hands on the wheel, ready to take over. In fact, unfortunately, a lot of drivers are acting a little more like what Austin Russell described: reading texts, maybe a little nap, making a TikTok. Experts I’ve talked to over the years say this kind of mushy middle period is the most dangerous part of the transition to fully autonomous cars. Some other carmakers have pulled back from this approach. Tesla, of course, has forged aggressively ahead. Now that Luminar and Volvo are pushing forward, too, we’ll see what it does to the landscape in the next couple of years. 

I interviewed Russell back in 2017 when he was only 23 years old. When his company goes public, he’ll be among the youngest CEOs of a publicly traded company. He dropped out of Stanford with a $100,000 grant from Peter Thiel, the controversial venture capitalist who co-founded Palantir. And as it happens, Palantir and Luminar are both going public in nontraditional ways. Palantir is doing a direct listing, which basically lets a company go public without creating any new shares and less scrutiny from the SEC. Luminar, on the other hand, is merging with what’s known as a special purpose acquisition company, a company that’s already public but doesn’t do anything — except exist to buy another company and take it public. It is, again, faster and requires less involvement or, some might say, oversight, from the SEC. 

Russell will not only be a super young CEO, but he’ll also hold about 83% of the voting power at Luminar. So investors better hope he’s right.

Correction (Sept. 29, 2020): A previous version of this article and podcast misstated Peter Thiel’s position at Palantir. They have been corrected.