Stefan Seltz-Axmacher founded Starsky Robotics with the goal of the early adoption of Level 5 autonomous systems, moving freight without any human in the vehicle at all. But Starsky shut down in March after failing to secure additional needed funding. HDT Equipment Editor Jim Park recently spoke to Seltz-Axmacher about what happened and why he believes adoption of true self-driving truck technology may be harder than we think.
HDT: So tell us about what you were trying to do with Starksy.
Seltz-Axmacher: Our view was that the primary challenge in North American logistics is that it's really, really hard to get someone to spend a month at a time in a truck. So what we were building was a piece of technology that let trucks drive without a person in them, and that was by making them remote control for the first and last mile and autonomous on the highway. And we went pretty far down that route. We actually last June became the first ever company to test a fully unmanned truck – no person in the vehicle at all on a public highway. But unfortunately we weren't able to raise our next round of funding. And I've been in the process of winding the company down for the last couple of months.
HDT: That's a regrettable conclusion to it all.
Seltz-Axmacher: Yeah, it is. Especially because I'm more and more confident that our specific approach was right. That approach is not just making those trucks remote control, but also by being the operator. And the reason why I'm so confident that is right is that modern robots are actually fairly bad and it's really hard to make them work in all conditions. As the operator, we could have made them only need to work in the conditions that were easiest for the robotics to work in.
HDT: In other words, pretty much straight down an interstate highway with not too much exterior interference from side streets and shrubs and taxis and ladies pushing baby carriages.
Seltz-Axmacher: I would say even more than that. If a typical truck moves about 450 miles per day, at 55, 65 miles an hour on average, that means they're really only driving six and a half to eight hours per day. If you have 24 hours in the day, that standard of service means that the early versions of these systems can be pulled over and waiting in rest stops 10 to 15 hours a day to wait out conditions that robots don't like. And those conditions can be everything from high traffic to rain to hail. All of those conditions are really, really hard for robots to handle.
HDT: What about five years from now; would those problems be worked out and truly driverless trucks be possible?
Seltz-Axmacher: People have been working on this for 10 years and have spent at least $70 billion on it. So far no one's worked out any of those problems. So far almost none of those folks have been able to take the person out of the vehicle on a regular basis with a wide-scale deployment. So whether that happens in five years, a year, or 50 years is really hard to say.
HDT: You know, that makes me feel pretty good as a former truck that it takes $70 billion to do what I used to be able to do practically in my sleep.
Seltz-Axmacher: What's interesting is, you and I are really impressed by computers, because computers can do stuff that we can't do. I'd be quite surprised if you could, off the top of your head, tell me what the square root of 827904.6 is. But your phone can do that super easily.
HDT: You're seeing the blue screen of death in my eyes. No, I couldn't possibly do it.
Seltz-Axmacher: [laughs] Because computers are so much better than us at a number of tasks that we find incredibly hard, we assume that it follows that they'll be really good also at tasks that we find really easy. And the inverse is actually true. It's a really hard robotics problem to make an arm that can pick up an object and hold it up. It's a really hard and mostly unsolved robotics problem to do something that we've been doing since we were toddlers, to walk. It turns out that humans are really good at exactly the type of things that robots are really bad at. And the opposite is also true, which is why so much money has been poured into doing something that, you know, frankly every 16-year-old can do more or less safely.
HDT: One of the things that I could never get my head around is that the technology seems to be available. Yourself and several others that are still out there [have showed off] a system that to my eye appears to work, and yet the industry isn't operating these things on a regular basis. They're still in testing. Why haven't we gone ahead with this full steam ahead at this stage?
Seltz-Axmacher: I'd say this is another one of the big problems with robots today. You are not alone in thinking that these things seem to work. In fact, many of the of the world's PhDs and smartest tech people in the world also think that these things seem to work. But just about every roboticist looks at these systems and knows exactly how little they actually work.
HDT: Okay. So what's the difference?
Seltz-Axmacher: There's a massive difference in robotics between being able to make something work once and being able to work a hundred out of a hundred times.
HDT: Well I think, if we were comparing performance of a human to a robot, the trucks would probably do a better job than most humans do right now.
Seltz-Axmacher: No, they wouldn't.
HDT: On a straight sort of a non-complicated piece of highway? I'm thinking like interstate 80 in Nevada or Arizona.
Seltz-Axmacher: If you made a track and that track had no traffic on it and the truck just had to drive on that track for a thousand hours, absolutely. The person loses every time. If you add in a mile of construction work here, people merging in and out there, it becomes a lot harder for the trucks to be equivalent. So far my company is the only company that has been so confident in the ability to drive on a straight piece of highway that we were willing to take the person out, and that was just for a single test. The amount of additional work we would have had to do to drive on that specific set of road, that specific 8-mile stretch, all day long every day, probably would've been another couple of years of work. Now when you start to think about the stuff that happens between the off ramp and the distribution center, that is a long ways away.
HDT: We have things like precision GPS now that can steer a vehicle down the road to within inches from one side of it to another. Are your trucks or trucks like yours equipped with that sort of technology or is it all quote unquote visual from the robot's point of view? LIDAR, radar, that sort of thing.
Seltz-Axmacher: There's plenty of people who are using lots of different sensors. High-accuracy GPS, high-definition maps, sensors that can tell you within a hundredths of a millimeter where are you are based on where the stoplight is and where the stop sign is if you've been there before. The real problem is all of the stuff that is not predictable, which is, human behavior, which is animal behavior. All of that stuff is what's really, really hard for robots to do.
The reason your calculator can tell you what the square root of that number I said earlier is, is because every time you type in that number and hit the square root sign, the answer is the same. When you're about to make a wide left turn on a road that is sometimes busy and a car is approaching from far off, it is really hard to predict what that car is going to do. Are they going to slam on their brakes? Are they going to pull over? Are they gonna speed up? Are they going to slow down and let you pass? Are they going to flash their brights at you to let to tell you to go. It's really hard to predict that, and given that a misdiagnosis can lead to a safety incident, it's really hard for teams to feel confident to deploy that type of system.
The approach that we took at Starsky was that we cared a lot about what we called consecutive zero-disengagement trips. So a zero dis-engagement trip was a trip where from the beginning of the trip till the end of the trip, the safety driver, the person sitting behind the wheel, never had to do anything. And if we did a large number of those trips in a row, where a large number of consecutives zero engagement trips, we could make a pretty good bet, a pretty good statistical bet, that on that next trip we would not need a safety driver in the vehicle.
HDT: How do regulators figure out how to address autonomous trucks?
Seltz-Axmacher: There are a number of states where it is already entirely legal for autonomous vehicle companies and autonomous truck companies to deploy systems on public roads without a person in the vehicle. What's been interesting and has particularly peeved me the last the last couple of years, is that a lot of the autonomous folks blame regulators and say, ‘Regulators are just never going to let this happen.’ When in actuality the reason we don't have autonomous cars is because the autonomous companies just aren't building something that works well enough yet.
HDT: You know, I've often said, and I, and I don't mean this in a disparaging way, that the tech community dreams up solutions to problems that we don't really have yet and then sets out to solve them, when we don't even maybe want them solved. So is this whole autonomous vehicle thing really just a pie in the sky right now?
Seltz-Axmacher: So I think it will actually happen. I think we do want this solved. We don't want 4,000 people to die every year in automotive accidents. The whole world economy and especially the U.S. economy would be changed meaningfully for the better if there was no longer a structural driver shortage. Like right now it's far cheaper to ship goods from China to Los Angeles than it is to, to ship goods from, let's say Ontario to Denver. That doesn't make sense. And autonomous trucks can change that. That being said, the technology just isn't quite there yet.
HDT: The trucking industry is sort of ambivalent on this or, or bipolar maybe – some really want them, some are dead set against them. When you were first approaching the industry, how did they respond to your ideas?
Seltz-Axmacher: A lot of people think that the trucking industry would hate this, and there's a long history of the trucking industry not being a quick adopter of technology or being ambivalent to Silicon Valley Wiz kids with fancy coffees and expensive jeans. But that hasn't actually been what I found. If you're going to build a tech product, you need to spend a lot of your time talking to the customer. So when I started Starsky obviously I was reading a lot about the trucking industry. I was diving in as deep as I could, but also worked to make sure I talked to some truck drivers, some small fleet managers, some big trucking executives. And what I found was that while they are not the typical purchasers of software, while they are not the customers that most of my tech entrepreneur friends get to sell to, the one thing that they do clearly want to buy is a solution to the truck driver shortage. This is a pants-on-fire problem for just about every fleet. And if we at Starsky were able to solve that problem, we would have no shortage of people who wanted to work with us.