Tesla said Wednesday that its new cars now have all the cameras, sensors, radar technology and computing power they need to truly drive themselves. That raises the prospect that hundreds of thousands of robotic cars could be on the streets far sooner than many thought possible (at least once production ramps up, software improvements are complete and the system is switched on.)
I spoke with automation expert Costa Samaras, an assistant professor of civil and environmental engineering at Carnegie Mellon University, about the implications for safety, the fiercely competitive autonomous car industry, and the future of backseat snoozing on cross-country trips. Here’s an edited transcript.
Michael Laris: Telsa said in its announcement that the cars they’re building have “the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.” Elon Musk emphasized that comparison. What’s the back up for that?
Costa Samaras: I don’t know. The standard that we like to think about is, it’s at least as safe as a human driver. In the policy and technical community, that’s where we want to be. He’s setting the appropriate goal. Now the challenge is proving that. You can prove it with simulations. Some of my colleagues at Rand had a paper that says proving safety in automation is going to take hundreds of millions, or even hundreds of billions, of road miles. That could take decades or even centuries. So there’s going to be this gray period where we’re going to be beta testing automation as it gets on the roadway with a hybrid of simulation and road tests. That’s not to say this is not an exciting development. We need to make huge leaps forward in road safety, and automation is the way to do it. The challenge is figuring out how to have that appropriate balance of existing safety and automation.
Q: On simulations, companies talk about driving millions and millions of simulated miles. Is that a good way to go? Tesla has also said they use modeling based on data they get from their current fleet of cars. They make what sounds like a good argument they can learn quickly that way. Is that true?
A: I’d be hesitant to say whether that’s true or not, without actually being there and looking at what’s happening. What I would be able to say is we will need a hybrid of simulation and deep learning across the fleet, as well as on-road miles to make this work. I’ve seen some companies propose and actually start doing simulations [where] the vehicle would be viewing a video that actually happened on the road, but they might put in another car or an obstacle or an animal that crosses the road. They can do that hundreds of thousands or millions of times. It’s kind of this simulated on-road experience for the vehicle. That’s the best option we have right now.
Q: A new Tesla video is pretty compelling. It starts with a little dig at regulators at the start, emphasizing that the driver’s only there for legal reasons – and that the car really is driving itself. There’s a crazy motorcyclist speeding past the more cautious robotic driver. The Rolling Stones’ “Paint It Black” blares as the car does its own thing. Which all seems pretty cool. So how much should people just go along with Elon Musk’s go-go enthusiasm, or should people really have their BS detector out and ready? What’s the balance between wonder at amazing technological advancements and taking a more cautious, trust-but-verify approach?
A: I think we can have both. In hardware, you want things to perform as advertised. If you think about the way we deal with airlines, we want to make sure that airplanes are safe every time they go in the air. For this, this is something where all these sensors, all these components, need to be as safe as the human driver. What that means is, what are the failure rates going to be? What are the maintenance cycles going to look like? What’s the redundancy that’s built in? If the computer has a command to brake and there’s a short or there’s some sort of mechanical problem, is there another path where the thing can brake?…These cars have to be safe every time they get on the road. Airlines take a break every so many hours and get checked and get maintained. Many states have state inspection programs, but some don’t. How we deal with this shifting of responsibility – from the driver to the car and ultimately to the manufacturers – is getting sorted out.
Q: There’s an undercurrent among some competitors, who quietly hint that Tesla is being reckless by pushing things out too fast. Is Tesla being reckless? Is Tesla a reckless driver?
A: There’s a balance. There’s an engineering set of challenges and there’s a business and marketing set of challenges. Sometimes they are in agreement and sometimes they’re not. The most important thing is delivering on performance. Sometimes in tech there’s a huge hype cycle and we get underwhelmed when the product is in front of us. I hope that Telsa can deliver. By announcing and being aggressive, maybe that spurs innovation across the entire industry faster. Books will be written about how this all played out. I think the danger is public perception right now. If there are high-profile crashes and failures and injuries and fatalities that look to be something where the technology is not ready, that could set the entire industry back. It’s not going to be fair, because there are 35,000 fatalities in the U.S. every year. That’s about equivalent to a fully-loaded 747 crash every week. There will be high-profile gray-zone boundary crashes as we transition to automation that will be in the public’s mind, whereas the 35,000 fatalities are largely out of the public’s mind. There’s a higher bar to meet in the first rollout. Even in the nuclear power industry, several high-profile accidents had an affect on a generation’s perception of an industry. That’s where I think everybody’s a little nervous right now. Again, this is the most exciting time to be in transportation since the Model T. What I try to do is highlight the positives but not forget about the negatives and try to work through all of them.
Q: Musk said there should be a fully autonomous test drive from LA to New York by the end of next year. Other companies have suggested they might start by running in a city here or there. Can Tesla really go nationwide? Is what Musk saying the same as saying, You can go from a rural Texas back road to a snowy Colorado interstate?
A: If you’re driving from LA to New York, which could be done right now with the technology, you need somebody behind the wheel. If you’re talking about driving a car by itself from LA to New York, that’s a whole other level of tech advancement. Because you need to insure that nobody will need to be behind there to take the wheel [and that] a machine can do the whole thing…We can do LA to New York right now, but I can’t sleep in the back the whole time. The thing I’d like to know from the industry is, when are there vehicles with nobody in them on all roads? That’s a different question than, this vehicle can navigate the road without the driver taking control again in most cases.
Q: You said the Rand paper talked about needing hundreds of millions, or even hundreds of billions, of miles of testing. Is this actual miles on the road? And before you get to what?
A: Even though there’s a lot of crashes on the road right now, with regular drivers, there’s also a whole bunch of miles. So the failure rate, the crash rate, the fatality rate, is small on the order of magnitude of the number of miles. We drive 3 trillion miles a year in the United States. Twelve zeros, right? In order to prove statistically that this new technology has that same rate, you need to have lots and lots of miles. This is just traditional reliability engineering: how many of these fail per ones that don’t fail? And in order to do that, you need lots and lots of observations, in this case driven miles on the street. So the highlight of that paper, which I thought was really elegant is: We’re never going be able to road-test our way to verification for safety for robotic cars. But that’s okay. We won’t have the time to do that. The subtext is, that’s fine. We’ll figure it out.
Q: With modeling or something else?
A: With modeling, with hybrid modeling and road testing. We’re going to muddle through. And that’s what we should do. As long as it’s in the zone of better than human drivers, we should be doing it.
Q: So you got off a plane this week and hear Tesla’s announcement, what’s your overall takeaway? How significant a moment is this? How does it fit in with everything else that’s happening?
A: This is really an exciting time for the way that people move around. We’re going to look back at these few years as really the start of the big transition to automation. That said, I’m an engineering technology analyst, and in that role I have to be objective and cold about all of these things, and that’s healthy. Making sure everything works as advertised is part of our role in the larger community.