At Tesla’s “Autonomy Day” event for investors on Monday, Elon Musk was full of trash talk for his competitors and their technology.
LIDAR, the light beam sensor that practically everyone views as an essential ingredient for self-driving cars, is “a fool’s errand,” according to Musk. What’s more, he went on, “anyone relying on LIDAR is doomed. Doomed. Expensive sensors that are unnecessary. It’s like having a whole bunch of expensive appendices… you’ll see.”
The simulation programs engineers use to run their virtual self-driving cars through millions of “edge case” scenarios to rack up billions of driving miles are good but not good enough, Musk said. “We have quite a good simulation, too, but it just does not capture the long tail of weird things that happen in the real world.”
High-precision GPS maps for self-driving cars are a “really bad idea,” according to Musk, resulting in a “system [that] becomes extremely brittle” by being too dependent and not being able to adapt.
Also, Tesla will have its Level 5 autonomous vehicles that can drive anywhere, under any condition, ready to go by 2020, according to Musk. No geographic constraints. And not just a handful of vehicles, either. “A year from now, we’ll have over a million cars with full self-driving, software, everything,” he said.
For those who have been watching Musk make wild predictions for over a decade, Monday’s event wasn’t all that surprising. After all, this is the guy who tried to turn his factory into an “alien dreadnought” of car-building robots. But the degree to which he seemed to enjoy knocking down the central pillars of the self-driving industry was truly astonishing, and it has many in the industry puzzling over how he can possibly deliver on his promises.
Musk has long argued that Tesla’s self-driving advantage comes from having a large fleet of vehicles — around 425,000 — already on the road. Those cars record situations and provide training data to improve the neural networks needed for self-driving cars. The company’s approach to autonomous vehicles is primarily focused on computer vision, or using cameras — just like humans — to recognize and understand the world.
Practically every other company trying to bring self-driving cars to the road — including Ford, Uber, Waymo, and GM Cruise — relies on a suite of sensors comprised of LIDAR, cameras, and radar. These companies argue that LIDAR can do things that cameras and radar cannot, while also providing overlapping capabilities to the things those sensors can do. These capabilities, known as redundancies, are extremely important for fully driverless vehicles as they provide an important backstop in the event of a failure.
Musk’s argument about LIDAR being useless with cameras because it replicates the visible light spectrum is “just wrong,” says Sam Abuelsamid, a senior analyst at Navigant, a technology consultancy. The human eye will respond to wavelengths from about 380 to 740 nanometers, while LIDAR can respond to the higher range of 905 to 1,550 nanometers. And new types of LIDAR, like the continuous beam FM sensor from Blackmore, can instantaneously measure velocity while also reducing computational latency, Abuelsamid notes. Layering sensors with different capabilities, rather than just relying on a purely vision-based system, is “ultimately a safer and more robust solution,” he said.
The Tesla CEO has been very outspoken and critical of LIDAR usage for autonomy, going as far as calling it “lame” at Monday’s event. “In cars, it’s freaking stupid,” Musk said. “It’s expensive and unnecessary. And as [Tesla AI director] Andrej [Karpathy] was saying, once you solve vision, it’s worthless. So you have expensive hardware that is worthless on the car.”
LIDAR can be incredibly expensive, and it’s a costly bet for most companies. They can make the job of selling self-driving cars to customers practically impossible thanks to the added costs. But companies that use LIDAR are already working on reducing those costs. Waymo began manufacturing its own cheaper LIDAR sensors in 2011. At the time, Waymo said it could lower the unit price from $75,000 for an off-the-shelf LIDAR sensor to just $7,500 with its own custom version. Last March, the company announced that it would begin selling its smallest LIDAR to third parties.
Investors who attended Tesla’s Autonomy Day event weren’t impressed by Musk’s anti-LIDAR tirade. “We continue to believe that Lidar is a good complement for cameras, radar, and ultrasonic sensors, and that the ultimate Level 5 winners are combining all of these sensors using sensor fusion,” Cowen’s Jeffrey Osborne wrote in a note on Tuesday. “Tesla’s rejection of the technology as a ‘fool’s errand’ due to the currently high additional expense is likely penny-wise but pound foolish, especially since Lidar prices continue to drop.”
Musk wasn’t as dismissive of simulation programs as he was of LIDAR, but he did claim that even the most sophisticated simulations fail to capture the ultimate “weirdness” of the real world. “If the simulation fully captured the real world, well, I mean that would prove that we’re living in a simulation, I think,” he said, harkening back to his comments at Recode’s Code Conference in 2016. “It doesn’t. I wish.”
His comments seemed targeted at Waymo, the company most see as having the lead in the race to make fully driverless cars. Waymo is big on simulation. It often touts how its Carcraft system has driven 7 billion miles in simulation, compared to 10 million miles on public roads as of last October. Musk argues that simulations are insufficiently weird to capture actual driving, which means Waymo won’t be able to catch up to all of the miles Tesla vehicles have driven with a simulation.
When asked for comment, a Waymo spokesperson referred to the company’s past statement on the importance of simulation: “One of the key advantages of simulation is that you can focus on the most interesting interactions — flashing yellow signals, wrong-way drivers, or nimble pedestrians and cyclists — rather than monotonous highway miles.”
Musk also criticized the practice of using high-definition maps to help guide self-driving cars through an environment. Before it deploys any vehicles in a city or town, Waymo first builds a detailed picture of that area and categorizes “interesting features,” like driveways, fire hydrants, and intersections. By knowing what the “permanent features of the road,” Waymo’s sensors can focus on moving objects like other vehicles and pedestrians.
According to Abuelsamid, the benefit of high-definition maps comes from being able to precisely localize the car in its environment by triangulating the distance from known objects. HD maps also contain data about the rules of the road, like the speed limit or which lanes you can turn from. This narrows down the area that must be scanned by the car’s perception system, and allows those sensors to focus on what’s most important: pedestrians and other cars on the road.
The biggest shock came when Musk stated unequivocally that Tesla would have “a million” driverless cars on the road by the end of 2020. The plan is for them to operate commercially in a ride-hailing network. When asked if these cars would be “Level 5 without a geofence,” meaning they could travel anywhere, under any conditions, without a human behind the wheel, Musk said yes.
Some analysts walked away with the impression that the robotaxi idea was underdeveloped, or “half-baked,” Cowen’s Osborne wrote. Some experts doubt that there can ever be a true Level 5 autonomous vehicle. As such, the vast majority of engineers in the AV industry are trying to perfect Level 4 autonomous driving, which requires zero input from a human driver, but only in a specific geographic area or under specific conditions (like good weather).
“For the foreseeable future (if not always) cars or other mobility devices will need to function under a set of engineering constraints,” says Bryan Reimer, a top research scientist at MIT’s Center for Transportation and Logistics. He noted that what the industry is working towards now is expanding the geographic domain for Level 4 vehicles, not the “unconstrained Level 5 problem.”
Reimer said that any engineering team working under a “reasonably adept regulatory authority” would not develop systems that drive in conditions that do not permit safe mobility, like blizzards or hurricanes. Musk suggested that his cars will be able to function under any scenario, no matter the danger.
Never mind the fact that the rest of the industry disagrees. At a tech conference last year, Waymo CEO John Krafcik said that “autonomy will always have constraints,” which was a sobering admission from the head of the leading driverless car company. Others have made similar comments about reining in expectations. “We overestimated the arrival of autonomous vehicles,” Ford CEO Jim Hackett said earlier this month.
It’s almost as if the entire industry was trying to make amends for all of the hype and cash that has been dumped in this space over the past few years. Everyone except for Musk, that is.
“While Tesla will certainly continue to introduce better assisted driving features over the coming years, true autonomy is a completely different game altogether and requires a completely different approach,” said Austin Russell, CEO of LIDAR maker Luminar. “There’s no longer a backup driver, hands on the wheel, taking over when things frequently go awry. No doubt it will get [twice, three times] or maybe even [10 times] better over the vehicle’s life cycle. But for autonomy, it has to be, quite literally, near perfect … to achieve safety greater than that of the average human driver.”
The argument could be made (and it has been made in the comments section of previous Verge articles about Tesla) that Waymo and other AV firms need to battle it out with upstarts like Musk in order to produce the best outcome for us, the customers. If Waymo is the only game in town, we’ll all probably pay more for rides and be subjected to a lot of forced Google tie-ins, ads, and data sharing about our movements. With Tesla (and Ford and GM) in the mix, Google will be forced to move faster and behave better.
Another way of looking at it, though, is that by bucking the rules of what is safe in self-driving cars and what isn’t, Musk is putting people in danger right now. According to a new poll, 71 percent of drivers across the world think they can buy a driverless car today. They cannot, of course, but they think they can because some carmakers are designing and marketing vehicles in such a way that drivers believe they can relinquish control. Tesla has been criticized for selling a “Full Self Driving” option that misrepresents what its vehicles are capable of doing.
Musk wants to gain a competitive edge by referring to Tesla as “full self-driving” vehicles, but he is fueling consumer confusion. And a confused driver can be worse than no driver at all.