Self-driving cars are a scam,” says George Hotz, the 28-year-old CEO of Comma.ai who rose to fame under his “geohot” hacker alias when just a teenager. We are sitting at the headquarters of his self-driving startup: a gorgeous, two-story California-style home in the tony Forrest Hills neighborhood of San Francisco — a big step up from their old digs in Potrero Hill. For over two years now, Hotz has been teasing a $1,000 after-market kit called Comma One that would let customers transform their dumb cars into smart, self-driving ones. But after a sternly worded letter from the US National Highway Traffic Safety Administration last year, Hotz abandoned those plans and instead published his autonomous driving code, “openpilot,” online for free.
When Hotz says this, we are minutes away from a test drive in a Honda Civic equipped with the latest version of his do-it-yourself semi-autonomous driving kit with driver monitoring, which is available today to download. But I’m a little confused: if self-driving cars are a scam, and he abandoned his plans to sell his Comma One after-market kit, what the hell is he selling?
Hotz seems to enjoy controversy. At age 17, geohot made a name in hacker circles as the first person to carrier unlock the iPhone. A few years later, he got in trouble with Sony for hacking the PlayStation 3 (the company sued and then later settled out of court). In 2015, he got in a fight with Elon Musk, after Musk reportedly tried to hire him, for claiming he could make a better version of Tesla’s semi-autonomous driving software Autopilot. Tesla called his claims “extremely unlikely.”
The flaw with Waymo and Tesla and all the companies working on autonomous driving, he says, is their desire to remove humans from the equation. “Every self-driving car on the road today is worse than a human,” he tells me. “Everyone, Waymo included. So with a human we believe these systems are safer than a human alone. And they certainly can be more convenient.”
The real money is in advanced driver assist systems (ADAS), much like Tesla’s Autopilot or Cadillac’s Super Cruise, Hotz says. That’s what sets Comma.ai apart from the majority of companies working on self-driving cars. In addition to being significantly larger and better funded than Hotz’s startup, most companies in this space plan to deploy their vehicles for commercial ride-hailing services. In Hotz’s view, autonomous ride-hailing will be too slow and inconvenient to have much of an impact on our transportation habits.
“The self-driving product is a worse Uber,” he says. “They already have this product, it’s called Uber, it works pretty good. The self-driving version will just take you 50 percent longer to get to your destination.”
So in order to compete with the larger companies, Hotz is going directly to consumers. Using three gadgets, Hotz says he can override the driver assist systems of any late model Toyota or Honda vehicle, and replace it with his own openpilot software. The three gadgets are a modified OnePlus smartphone, called the Eon, that’s like a souped-up dashcam; a dongle already for sale called the Panda, that allows external electronics to interface with a vehicle’s computer system; and the Giraffe, which enables Hotz to issue commands to the car via software by connecting to a vehicle’s CAN (controlled area network) bus. All together, these gadgets allow drivers to use Hotz’s version of lane-keeping and adaptive cruise control — all without having to touch the steering wheel.
Cadillac’s Super Cruise is the only system on the road today that offers hands-free driving, thanks to an infrared camera mounted on the steering column that monitors the driver’s eye movements. If the driver looks away from the road too long, the car will issue a series of haptic, visual, and audio alerts until the driver turns his or her gaze back. If the driver ignores the warnings, the car will disengage Super Cruise and the vehicle will eventually start to slow down. Super Cruise also only operates on 130,000+ miles of highways in the US and Canada that GM has Lidar mapped.
Comma.ai’s system is a scaled-down Super Cruise, without the mapping and mandatory geofencing. A camera embedded in the Eon monitors the driver’s head movements. If the driver looks down or away from the road for two seconds, a visual warning will appear on the Eon’s screen. After four seconds, the driver will hear an audible alert. And after six seconds, the system will disengage and begin to slow down. Unclick your seat belt or open the car door, and the system disengages.
Overreliance upon hands-free driving technologies can be dangerous, especially without a driver-monitoring system. Tesla has been criticized for not including a robust driver monitoring system to buttress Autopilot. And Tesla and Cadillac are luxury vehicles, affordable only to a small segment of the population. Other automakers have their own ADAS — Nissan ProPilot, Honda Sensing, etc — but the consensus among car and tech reporters is that none are as good as Autopilot and Super Cruise. Hotz wants Comma.ai’s system to be the best one on the market. “Our goal is to make our system better than Super Cruise by the end of the year,” Hotz says.
Climbing into the driver’s seat of Hotz’s silver Honda Civic, the first thing I noticed was the absence of a rear view mirror. The Eon, connected to the vehicle’s backup camera, is attached to the windshield in its place and provides a view from behind the car.
I’ve driven a Cadillac CT6 with Super Cruise and a Tesla Model S with Autopilot. Both are extraordinarily good driver assist systems. But when I’m driving southbound on Interstate 280, I realize they might have a viable competitor in Comma.ai. The car handles well at speeds of up to 75 mph, taking curves, passing underneath bulky overpasses, and braking seamlessly for merging vehicles.
After some adjustments, I don’t once have to touch the steering wheel or pedals, nor do we encounter anything that would make me feel like I needed to take control. All the same, I keep my fingers lightly resting on the steering wheel. I’ve written too many stories about accidents involving semi-autonomous cars to feel comfortable taking the chance.
At no point am I forced to take control of the vehicle, but then again, we’re driving in a geofenced area, which Hotz says was just for the purpose of demonstrating the technology to the press. “The press love it when things mess up,” he says. Of course, this opens Hotz up to censure from his detractors: Tesla previously criticized the hacker for only showing off his technology in a “limited demo on a known stretch of road.”
Somewhere outside Daly City, we pull over and switch seats. Hotz turns off the geofencing, and we leave the interstate and begin driving north on Junipero Serra Boulevard, a two-way, four-lane road with all the variables like bicyclists, train tracks, and complex intersections that make most self-driving operators sweat bullets. As long as there’s a vehicle ahead of us, the Honda decelerates and brakes at traffic stops. But almost as soon as we leave the geofencing, the car’s system disengages. Moreover, Hotz explains that he needs to take control through several intersections. “Whenever you see something where there could be ambiguity, put your hands on the wheel,” he says.
As the developer, Hotz obviously knows which situations are ambiguous enough to take control, but the same can’t be said for others using his technology. When it boots up, Comma.ai’s system warns the driver to be ready to take control at any moment. Drivers will need to learn its idiosyncrasies, which Hotz argues is the same for any modern, high-tech car on the road today.
I ask him about the amount of lane markings, road quality, and signage required for the system to work. “The bare minimum is you can engage it everywhere,” Hotz says impishly, “and it works to varying degrees of goodness.” When the final product eventually ships, though, Hotz says he envisions there being a geofence, much like Super Cruise. Or not? It’s hard to pin him down because it’s not clear that Hotz has really thought all of this through. He’s even considering including an “expert mode” without geofencing that can be unlocked after driving a certain number of miles. Hotz is a video game buff.
California’s DMV requires all companies testing self-driving cars to report how many times safety operators were forced to take control of its vehicles. Every year, when these reports come out, there’s generally a lot of hoopla because these are the only metrics we have to analyze how well each company’s tech performs. Comma.ai does not have a permit to test self-driving cars, so it is not required to report its disengagements to the DMV.
Still, the company wants to cut down on disengagements through its new “Explorer” tool, which allows the company’s community of users to crowdsource disengagements and fix them in future versions of the software. The idea is that there are common scenarios when the system disengages that can be fixed through software patches. Users upload data from their drives to Explorer and then select from a list of canned reasons for the disengagements. The reasons range from “turn too sharp” and “took an exit,” to “arbitrary or accidental” and “I needed to take over for safety.” Hotz’s team then feeds the disengagement data into Comma.ai’s machine learning software and uses it to make improvements in later versions of openpilot. Comma.ai will tackle disengagements for safety concerns first, Hotz says.
“This is not some altruistic you’re-making-it-better-for-everybody kind of thing,” he says. “Really it’s making it better for you.”
Hotz drives us back to the Forrest Hills house, where the Comma.ai team is working to get shipments of the company’s trio of gadgets to buyers around the globe. Hotz has over two dozen 3D printers in the garage building plastic casings for the Eon, Panda, and Giraffes. Stacks of padded envelopes sit by the door, waiting to picked up by the postal service. A framed poster of a Tesla Model S with a big red comma superimposed over it hangs on the wall. Also mounted on the wall is a vanity license plate that reads “FUELON,” with a pink heart covering the “F.” Ethernet cables are strung along the ceiling and staircase. Everyone walks around in their socks. It feels like the tidiest fraternity house I’ve ever seen.
The goal is still Level 4 autonomous driving, which most define as a fully driverless vehicle within a specific set of circumstances, but Hotz defines as shifting the liability from the driver to the manufacturer in the event of an accident. That’s why he is calling today’s announcement about the new version of openpilot with driver monitoring the “halfway point.”
“We’re halfway to a consumer product,” he says. “The hardware is all available for sale in our shop. And the software will all be on our GitHub by the time this article goes live. Every line of code you see running in this car is open source.”
More than Hotz’s bluster, the open source software really does set Comma.ai apart. And the startup has inspired others, such as Neodriven, a small company that’s selling a modified version of the plans that Hotz open sourced. Though Comma.ai only has 200 active users right now, those users are enthusiastic about Hotz’s work, posting unboxing and installation videos to YouTube and reviewing his products online. One YouTuber, Virtually Chris, posts videos where he puts Hotz’s system through its paces. “Really impressive,” he called the latest version of openpilot.
But open source does not translate into a viable revenue stream. Hotz says he hopes this latest round of press will cause those user numbers to grow. “I’m a businessman,” he says. “I want to make money.”
Last year, Hotz shut down his Comma One project after a letter from NHTSA expressed concern that the product “could put the safety of [Comma.ai’s] customers and other road users at risk.” The federal agency didn’t expressly demand or request that Hotz shut down the project. Rather, the government agency “strongly” encouraged Hotz to either delay selling or deploying Comma One on public roads “until [Comma.ai] can ensure it is safe.”
The Eon, Panda, and Giraffe can all be purchased online for around $1,000. (The Panda is even available on Amazon for $99.) But the software that enables the semi-autonomous driving is free to download. Hotz says this allows him to sidestep the regulatory issue, though it’s unclear whether NHTSA would agree. “We aren’t selling any products that control a car,” he says. “We are giving away free software, and software is speech.” (A spokesperson for NHTSA did not respond to a request for comment.)
Hotz is careful not to use the same rhetoric you might hear from other purveyors of autonomous driving technology. He doesn’t talk about safety as much as he does convenience and coolness. I ask him if he’s still interested in selling self-driving cars. “Regulatory agencies like to yell at me when I say that, so no, of course this isn’t a self-driving car,” he says, gesticulating wildly from his desk chair.
“Do I personally feel it is?” Hotz says. “Well, you know. Ask me when you’re not a reporter.”