Motor City Blogman

Self-Driving Teslas: Is Artificial Intelligence Intelligent Enough for Autonomous Driving?

Gaining a better understanding of Elon Musk's Autopilot plans.

We have reached an odd sort of moment in the development of the autonomous car in which skeptics have assured themselves it never will happen while proponents of the technology think of it as a done deal, and neither opinion doesn’t seem that far off from reality or even each other. In April, when CEO Elon Musk held court on the latest of his entertainingly controversial Tesla quarterly report analysts’ conference calls, he outlined a program that would allow his customers to loan out their Models 3, S, and X when not in personal use. And by the way, these cars would drive themselves to the next ride-hailing consumer.

Back in Dearborn and Detroit, despite investing $1 billion in artificial-intelligence company Argo AI, the Ford Motor Company is lowering expectations for the speed of its autonomous-technology development. About the same time Musk announced that Tesla’s new onboard computers make its current production models capable of Level 4 to 5 self-driving, Ford CEO Jim Hackett told the Detroit Economic Club that the industry has “overestimated” autonomous-car development. Ford’s much-ballyhooed Level 4 self-driving car coming in 2021 would be, well, just that: an SAE Level 4 autonomous car while operating within a heavily geofenced environment. That means the Level 4 Fords will pick up and drop off passengers with no driver behind the wheel, but only on geofenced roads and streets.

Meanwhile, after more than a decade developing and testing self-driving Lexus RXes around Silicon Valley and Phoenix and other warm-weather neighborhoods, Waymo (formerly Google Auto) has announced plans to assemble autonomous versions of Jaguar I-Paces (non-autonomous version pictured above) and Chrysler Pacificas in a new Detroit factory.

More recently, Nissan announced ProPilot 2.0, in which Nissan Skylines (the home-market version of the Infiniti Q50) can be operated hands-free along Japan’s network of highways. The Nissan Skylines will use 3-D mapping of the country in combination with driver-alertness monitoring, seven cameras, five radar sensors, and 12 sonar sensors (but like Tesla, no lidar) to allow hands-free driving from on-ramp to off-ramp. Japan’s network of highways relies heavily on automated toll booths, which helped facilitate the 3-D mapping, and the automakers considers this version of ProPilot (with the word “Assist” dropped from the name, in this case) to be SAE Level 2 autonomy, more like Cadillac Super Cruise and less like Tesla Autopilot.

To those of us witnessing autonomous development from an auto-manufacturer viewpoint rather than a tech-disrupter viewpoint, Nissan ProPilot 2.0 appears far more measured and responsible than Tesla’s headlong rush into full-ish (Level 4, geofenced in cities like San Francisco and New York, at least) self-driving capability. To Gary Silberg, partner in charge of automotive at professional services and big-four auditing firm KPMG, this is an example of “The Clockspeed Dilemma,” based on Albert Einstein’s Theory of Relativity. To Nissan or General Motors, five years of autonomous development goes by very quickly, Silberg would posit, while five years of autonomous development at Tesla is a lifetime, especially when you have hundreds of thousands of rolling computers/cameras collecting road data.

Musk said at his company’s “Autonomy Day” just before that April quarterly report call; “How could it be that Tesla, who has never designed a chip before, would design the best self-driving chip in the world? But that is objectively what occurred. And not ‘best’ by a small margin, but by a huge margin.”

His team then took fawning investors on a proscribed route around the San Francisco Bay Area in self-driving Teslas. Never mind that these routes were pre-mapped, much like those required by Nissan’s ProPilot and Cadillac’s Super Cruise.

Silberg first explained Tesla’s reliance on real-world data from its customers to a group of Silicon Valley skeptics at the Automotive Press Association in Detroit, and I spoke with him again a week or so later.

“One of the huge advantages is that Tesla has a computer on wheels already, over 400,000 of them according to Elon, that are on the road with the sensory data that allows them to teach the neural nets from their camera technology to the radar to the ultrasonics. The case that he makes intellectually is consistent with deep learning and artificial intelligence in that, if you have the data and you can use it and leverage it, it is invaluable.”

These data from its drivers gives Tesla far more road-test experience than all of Google/Waymo’s years of on-the-road testing, Silberg says. Artificial intelligence can learn anything, including how to identify what is in the pictures taken by its cameras to the point it can distinguish between a car or truck and a pedestrian, a dog, a cat, a rabbit, a motorcyclist, a bicyclist, or virtually anything else that might appear on the road, including a bike attached to the back of an SUV. Tesla has about 425,000 vehicles on the road globally (with about 1 million promised worldwide by the end of this year based on Musk’s projections for calendar-2019 sales) that constantly send data back to the automaker showing all sorts of driving conditions and encounters, including “corner cases,” the low-probability, high-impact events that are nearly impossible to predict in a simulation.

“It is a big advantage to actually see corner cases versus to try to create or imagine corner cases that you would just simulate. That to me is (Musk’s) greatest strength and argument, which I think is very compelling.”

You have to separate Musk himself from Tesla’s efforts, Silberg says, acknowledging the mercurial CEO’s propensity for exaggeratedly quick development deadlines. Silberg harbors no delusions that Tesla’s rollout of Level 4 or 5 autonomy will happen nearly as soon as its CEO claims, but he understands Musk’s impatience with working in more traditional ways to develop and test autonomous vehicles.

“If you’re GM and that’s your position in the marketplace,” he says of the automaker’s mapping and testing, “you have to have your own strategy because you don’t have enough vehicles out there. I’m not saying they didn’t make the right decisions. Tesla has that first-mover advantage that all its cars are computers on wheels. GM’s cars are not all computers on wheels.”

Call me unwilling to embrace the new technologies, but I’m more comfortable with the Luddite automakers’ belt-and-suspenders method of developing autonomous vehicles. As a sort of side note, I don’t appreciate the notion that the car I’m driving is sending its maker data of everywhere I’ve driven and of everything I’ve seen along the way. But Silberg’s explanation of “The Clockspeed Dilemma” and the way a company like Tesla can teach its cars’ computers how to recognize virtually anything on the road has helped my understanding of how tech “disrupters” can think in a completely different way about the same problems traditional manufacturers are trying to solve, whether it helps said tech disrupter’s stock value, or in this case . . . not.