Elon Musk Reportedly Rejected Driver-Monitoring for Tesla Autopilot—But Why?
Tesla needs to install a driver-monitoring system before it’s forced to—or Euro NCAP kicks in.
For all the genius that went into creating (and marketing) Tesla Autopilot, it's been missing a critical safety feature since day one: a driver-monitoring system. Why the blind spot? According to Tim Higgins at The Wall Street Journal, the answer might actually have been that Elon Musk knew how to solve Tesla's Autopilot issue, but chose not to.
Musk responded to the WSJ story within hours with the following tweet:
What is Elon Musk talking about? Cadillac SuperCruise, the only other system currently competitive with Tesla Autopilot, has a wonderful driver monitoring system that's available on the CT6 right now.
So what was Elon thinking in 2015 when he reportedly rejected a driver-monitoring system (DMS)? And what was he thinking when he wrote that tweet?
I've long been a fan of Tesla's innovations, but lost in the debate over Tesla Autopilot safety has been any real discussion over what it really is, and how a DMS would solve its shortcomings. Tesla Autopilot isn't autonomous, semi-autonomous, self-driving, partly self-driving or driverless. It's not a technology itself as much as a brand name for a suite of semi-automated features including radar-based cruise control (which Tesla calls TACC) and a lane-keeping assistance system that Tesla calls Autosteer. These are convenience features, and there is no evidence yet that they improve safety.
(The most popular safety figure attributed to Tesla, its 40% reduction in crashes, is almost certainly attributable to automatic emergency braking, common on most luxury cars and active on all Teslas even in the absence of Autopilot.)
Since its October 2015 launch, every generation of Autopilot has allowed limited hands-off operation. It can be engaged almost anywhere but won't stay engaged unless conditions are good, if not excellent. Neither its radar nor camera(s) are infallible, nor is the software interpreting what they see. Autopilot may disengage anywhere, at any time, which means the driver, who remains legally responsible at all times, should keep his eyes on the road and his hands on or near the wheel.
But no one ever listens, of course, because the "Autopilot" brand is too easily conflated with aviation autopilots, which are conceptually similar but functionally different, and require no less than one highly trained human at the controls. Most cars have none.
Every single Autopilot-related crash can be attributed to a lack of training—and, in my opinion, the absence of a DMS.
Commercial airliners have always the equivalent of a DMS; it's called a "second pilot." Even with the benefit of time, altitude, and two (if not three) pilots in the cockpit, the world's most advanced planes sometimes go down.
Which is why allowing cars with any hands-off operation—anywhere, anytime—without real driver monitoring and/or additional training, is insane.
Unless or until we raise drivers licensing standards, only a DMS can solve this problem, to the extent that it can be "solved" at all. Until zero-human input (what I call geotonomous) systems are commercially available, any semi-automated system with a steering wheel that allows any hands-off use is only as safe as the human "on-the-loop". At the very least, that requires drivers to be paying attention. Tens of thousands of people are killed every year due to distracted driving without the benefit of semi-automated driving systems; give human beings the slightest reason to pay less attention, and history suggests they will do just that.
Tesla has attempted to mitigate the absence of a DMS in two ways: 1) shortening the hands-off intervals; and 2) increasing the volume and visibility of hands-off warnings. Neither of these force a driver's eyes back to the road, an directive without which there is no possibility of a driver taking safe control when Autopilot disengages.
What are the options for a DMS?
There are are two distinct types of DMS, which serve different purposes: a control DMS keeps yours hands on the wheel, and an awareness DMS keeps your eyes on the road.
How to keep people's hands on the steering wheel? Two cheap systems exist right now: 1) a torque sensor that measures steering input; and 2) a capacitive sensor on the wheel itself.
A torque sensor is basically free to automakers, because it's built into any power steering system. It's easily fooled by fruits, vegetables, and water bottles jerry-rigged to the wheel, which means it's not a true DMS but a way to notify the system of your continued existence.
A capacitive sensor is basically free, if you've got a heated steering wheel. If not, the necessary coil is under ten bucks. You'll need some code, but no more than you need for a torque sensor. You can't fool this with fruits or bottles, which means this is what you want for your control DMS.
Guess which one Tesla used:
Musk can say skipping eyetracking wasn't about cost, but why skip a capacitive sensor? There's no question they work, so the only reason not to install one is cost.
What about an awareness DMS? As I explained in The Half Life Of Danger: The Truth Behind The Tesla Model X Crash, having your eyes on the road is more valuable than having your hand on the wheel:
So how to keep people's eyes on the road? You need a camera on the driver. Cadillac's excellent SuperCruise DMS camera is made by the Australian company Seeing Machines. The camera is on sale right now for $261. The wonderful light-up steering wheel upgrade, which includes the necessary infrared lamps, is on sale for $711.77, or $367 more than a standard wheel. That's a bit over $630 in hardware costs, plus whatever Seeing Machines charges GM.
Alternatives? Numerous cameras under $100 can run software from Affectiva, a company that does emotion recognition analysis that looks like this:
Or maybe Tesla should call Smart Eye AB in Sweden, or FotoNation in San Jose, not far from Tesla HQ. The vendor list is growing, and with good reason: Euro NCAP, the European voluntary vehicle safety system, made driver-monitoring systems a primary safety standard by 2020. In other words, any car that wants a five-star rating will have to have one. What that looks like hasn't been defined, but I'd bet it's a hands-and-eyes system. Which means Tesla better already have something better in the works, or the company will likely have to say goodbye to its current five-star safety ratings less than two years from now.
So what was Elon thinking back in 2015, if in fact he vetoed a DMS as suggested?
How do you say Minimum Viable Product in Fremont-speak? Autopilot. Cost. Time to market. Musk could have added capacitive touch for a few dollars a car, but didn't.
What about that "ineffective technology and overly sensitive sensors that would beep too often" part? Seeing Machines had solved it by that time, and Cadillac designed a beautiful system around that company's hardware. Musk wanted to be first to market with a semi-automated system; GM was willing to wait. Autopilot was released two years before Supercruise, and that extra time in development shows.
That Tesla Autopilot was ever any good at all is a testament to Musk's genius; that it's missing a critical piece of safety hardware is a testament to his impatience. Software is easy to upgrade. Hardware, not so much.
Why doesn't Tesla just add a real DMS now? They certainly could. True, it would be an admission that that current hardware is inadequate, but it would be wise to get ahead of the problem because there's no way around this omission, no matter how much people enjoy Autopilot (myself included). The more Teslas are sold, the more the familiar accidents will happen over and over. Even if the Autopilot brand goes away, the cat's out of the bag. People will always use technology as they wish in addition to or instead of as intended—which is why I predict camera-based DMS will be mandated for all hands-off systems.
Besides, that Euro NCAP DMS standard is coming in 2020.
And what of Musk's tweet attacking the WSJ story? Probably no more than a backhand. Musk isn't stupid. He knows DMS hardware has leapfrogged Tesla's system. My guess is that Musk is betting everything on Fleet Learning, neural nets, and artificial intelligence—praying Autopilot becomes Enhanced Autopilot, then evolves into Full Self-Driving before he's forced to admit a mistake. (He also thinks that will all happen without Lidar, and with the camera technology currently being installed on his cars.)
That's a lot of moving parts. There's a reason there's so much turnover in the Autopilot division, which is rarely a good sign.
Despite what his fans want to believe, Musk is a carmaker like any other, and that industry has a long history of cutting corners. I want Musk to succeed, because the whole industry benefits from innovation, but sometimes the light that shines too bright burns half as long.
Musk loves to surprise us. Maybe he'll make heated steering standard on all Teslas, declare capacitive touch has arrived, and make it a win-win for everyone. Hands-on will never be as safe as eyes-on, but it sure does help.
But who knows? The Tesla Model 3 has a tiny cabin-facing camera above the center rear-view mirror. I've yet to see anyone pull one out and determine whether it could be used for driver monitoring.
I doubt it, but anything is possible.
Alex Roy—Founder of the Human Driving Association, Editor-at-Large at The Drive, Host of The Autonocast, co-host of /DRIVE on NBC Sports and author of The Driver—has set numerous endurance driving records, including the infamous Cannonball Run record. You can follow him on Facebook, Twitter and Instagram.