There is a race going on right now that stretches from Silicon Valley to Detroit and vice versa: who can make a car without a driver that behaves better than a human driver? It is a much more difficult task than it seemed until a few years ago, because human drivers know a lot, not only about their cars, but about how people behave on the road when they are behind the wheel. To reach that same kind of understanding, computerized cars need a lot of information. And the two companies with the largest amount of data at this time are Tesla and Waymo.
Both Tesla and Waymo are trying to collect and process enough data to create a car that can handle itself. And they are approaching those problems in very different ways. Tesla is taking advantage of the hundreds of thousands of cars he has on the road by collecting real-world data on how those vehicles work (and how they might work) with the autopilot, his current semi-autonomous system. Waymo, which started as Google's stand-alone automobile project, uses powerful computer simulations and feeds what it learns from them in a smaller, real-world fleet.
It is possible, and advocates say without a doubt, that autonomous driving technology would reduce the number of annual deaths in the United States. UU As a result of car crashes, an amazing 40,000 people. But there is also a great financial incentive to apply all this technology based on data to the road as quickly as possible. Intel believes that autonomous vehicles could generate $ 800 billion per year in revenues in 2030 and $ 7 billion per year in 2050. Last summer, Morgan Stanley analyst Adam Jonas said in a note that the data could be more valuable for Prove that something like Model 3. "There is only one market large enough to boost the value of the action to the levels of Elon Musk's aspirations: that of miles, data and content," he wrote in June.
Fixed image of a Tesla video that shows that the autopilot is in action.
It is difficult to pinpoint exactly how many miles of data Tesla has obtained from Autopilot, because the company does not make many public statements about it. In 2016, the then head of Autopilot told a crowd at the MIT conference that Tesla had logged 780 million miles of data, with 100 million of those miles while the autopilot was "at least in partial control" according to IEEE Spectrum . Later that summer, Musk said that Tesla was collecting "just over 3 million miles [of data] per day." However, since last July, the total number of miles traveled by the fleet had increased to 5 billion. As Tesla sells more cars, the amount of data that can be collected increases exponentially.
Tesla customers have driven billions of miles in the real world
Not all those miles come from the autopilot, and the autopilot is still only a semi-autonomous feature. But Tesla also collects data on how the autopilot would handle different driving scenarios, even when the feature is not used. The Tesla cars can register instances in which the Autopilot software would have performed an action, and those data will finally be reloaded to Tesla. This so-called "shadow mode" of collection means that Tesla could be simulating autopilot data on many of those billions of miles that are handled.
The only other company that works with similar amounts of data is Waymo, which announced earlier this year that it has simulated 5 billion miles of autonomous driving. The company also said it has cut 5 million self-imposed miles on public roads. That's more than basically any other company that tests self-contained vehicles combined, if the recent reporting figures in the state of California-the largest seedbed of autonomous tests so far-are any indication.
Waymo is limited by the fact that it is only collecting real-world data through a fleet of 500 to 600 Pacifica self-driving minivans. Tesla has more than 300,000 vehicles on the road around the world, and those cars are navigating in much more diverse environments than Waymo, which is currently only found in Texas, California, Michigan, Arizona and Georgia. But Tesla is only learning from those miles in the real world, because even when the autopilot is activated, the current version is only semi-autonomous.
This balance will also change. Waymo plans to add "thousands" more Chrysler minivans to its fleet starting at the end of this year. And recently announced an association with Jaguar Land Rover to develop a fully autonomous version of the all-electric I-Pace SUV from scratch. Waymo says it will add up to 20,000 of these to its fleet in the coming years, and will be able to handle a volume of 1 million trips per day once all those cars are on the road.
Until then, Waymo relies heavily on their simulations, and computers can not always come with all the strange real-world scenarios. That's why it matters that Tesla is now leading in real miles, says analyst Tasha Keeney, who covers the company for Ark Invest. "I feel everyone agrees that Waymo's technology is the best at the moment, but I think a lot of people are underestimating the power of Tesla's data set," he says.
Photo of Amelia Holowaty Krales / The Verge
TYPE OF DATA
These two companies not only collect data at different scales, but also collect different data. Waymo's self-directed minivans use three different types of LIDAR sensors, five radar sensors and eight cameras. Tesla cars are also equipped with a lot of equipment: eight cameras, 12 ultrasonic sensors and a forward-facing radar.
But Tesla does not use LIDAR. LIDAR looks a lot like radar, but instead of radio waves, it sends out millions of laser light signals per second and measures how long they take to recover. This makes it possible to create a high resolution image of the environment of a car, and in all directions if it is placed in the right place (such as the top of a car). It maintains this precision even in the dark, since the sensors are its own source of light. That's important because cameras are worse in the dark, and radar and ultrasound are not as accurate.
LIDAR can be expensive and bulky, and also involves moving mechanical parts (for now, at least). Recently, Musk called the technology a "crutch", and argued that while it makes things easier in the short term, companies will have to dominate camera-based systems to keep costs down.
A large part of the industry agrees that LIDAR is necessary, but Musk does not agree
If Tesla can develop autonomous cars without that technology, Keeney says it would be a big advantage. "It's a riskier strategy, but in the end it could be profitable," he explains. "If Tesla solves [self-driving cars without LIDAR] all the others are going to kick themselves".
That is a great "yes". Without LIDAR data, Tesla may be at a disadvantage, according to Raj Rajkumar, co-director of the autonomous and connected conducted research laboratory sponsored by General Motors at Carnegie Mellon University. (CMU is a school so famous for its robotics skills that Uber picked up dozens of employees in 2015.)
LIDAR is seen by many in the industry as an essential tool to create cars that can handle themselves, and Rajkumar says there is great skepticism about Tesla's approach. "We do not think the hardware is enough to do that, and I do not think Tesla is particularly close to getting to [fully] operation without a driver," he says.
Tesla declined to comment on what data is being collected from what sensors, or the quality of that data. It could be all the video of the car, only some cameras at certain times (such as collisions) or data from ultrasonic sensors without video. And, Rajkumar says, it's also not clear if it's the full frame rate video or something with less fidelity.
Keeney agrees. "The Waymo data set is much more detailed just because they are using LIDAR, which gets a lot more information than it would get from the cameras alone," he says.
Collecting data is one thing. But even Musk has noticed that processing data is also a difficult task. "Actually, it's a big challenge to process that data, and then train against that data, and make the vehicle learn effectively from the data, because it's just a lot," Musk said in a profit call on last summer.
Waymo, comparatively, sounds more confident about its simulations. The company recreates complete computer models of the cities in which it is testing and sends 25,000 "automatic driving virtual cars" through them every day, according to a report in The Atlantic last summer.
This helps Waymo create a feedback loop adjusted by recreating real-world driving data on the computer, where "thousands of variations" of a scenario can be executed. Then, the information is downloaded back to the Waymo test cars. Waymo has also built a dedicated test facility in California, where you can build characteristics of particular streets or scenario scenarios that seem to give your vehicles the most problems.
Waymo has a more obvious loop between its simulations and its real-world test fleets
This closed cycle, says Rajkumar, "has come at the cost of incredible investments, resources, time and effort, which Waymo, of course, has a lot due to its parent company." He says Tesla would find it difficult to match this. . "Tesla would have to spend a lot more on that, and go through a highly labor-intensive process."
In his second "master plan" for Tesla, published two years ago, Musk said he believed it would take about 6 billion miles to obtain the "global regulatory approval" of true self-driving technology. It is likely that Tesla has already exceeded that mark in real miles, and yet its cars still can not drive on their own. A test drive of a Tesla driving from Los Angeles to New York was delayed, which was supposed to take place in 2017, and the target for the release of the final version of Autopilot continues to advance.
Meanwhile, Waymo is close to that figure of 6 billion miles on the simulation side, and the company is accumulating virtual miles faster than ever, with thousands more test cars waiting in the wings. He plans to launch a commercial drive program with his automatic driving minivans later this year, something he's already testing in Arizona, which could further reinforce that data feedback loop.
Photo of Vjeran Pavic / The Verge
Tesla and Waymo are two of the most advanced companies that test this technology, but they are not alone. One of the most visible competitors in this space has been Uber. Compared to Tesla and Waymo, Uber took a more casual approach with its autonomous driving tests, which is typical of the company that has epitomized Silicon Valley's "move fast and break things" slogan.
After starting tests in Pittsburgh in 2016, Uber placed the first versions of their modified semi-autonomous Volvos on the streets of San Francisco without obtaining the necessary state permits. When the company was arrested, they transferred the evidence to Arizona. Uber eventually agreed to the basic requirements of California, but his remains with lawmakers there put the company behind competitors like Waymo in the real world driven by miles.
Once it was set up with test fleets in three states, Uber quickly clicked on miles. It reached 2 million miles nationwide in November 2017, according to The New York Times. However, it is not clear how many miles Uber has simulated, and the quality of its technology has been questioned after one of its test cars killed a pedestrian in Arizona in March. Uber general manager Dara Khosrowshahi has said that the company remains "absolutely committed" to the program, but its testing efforts remain suspended for the time being.
The only other company that performs quality work similar to Waymo or Tesla when it comes to driverless cars, says Keeney, is one more outmoded: General Motors. GM has been developing self-conducting Bolt EVs with the help of a company it acquired, called Cruise Automation, and recently announced plans to test its own limited commercial self-driving service in 2019.
GM is designing the all-electric Chevy Bolts without steering wheel or pedals, and will launch a commercial version with the Bolts that has been modernized with Cruise Automation technology in 2019. Image: GM
GM is following Waymo's steps in generating and processing the data needed to teach cars how to drive themselves with small test fleets. But Keeney believes that GM's strength is its scale of production. "Waymo has this deal with Jaguar, and that may become something in the future, but they are not really producing the cars in the company, I think there is an advantage to having a vertical strategy," he says. "With a set of autonomous sensors, when you build it from scratch, you have a better idea of how production should be and how you can optimize everything."
GM, like Tesla, also has a semiautonomous product in cars for customers that are on the way at the moment. But that product, Super Cruise, is limited to a Cadillac model, and there is no sign of it spreading to other models in the short term.
In Keeney's eyes, that's another missed opportunity. "That's what they're losing, and that's what other automakers lack," she says. "Why has no one put sensors on their customers' cars that collect data like Tesla?"
WHAT IS GOOD SIMULATION, ANYWAY?
There is a dark horse in the race: Nvidia. It may not be accumulating billions of miles that Tesla and Waymo boast about, but Nvidia's technology is being used by hundreds of companies, including Tesla, in the autonomous driving space. Last month, Nvidia began selling what it calls "Drive Constellation", which is essentially a simulator already ready for the self-driving projects of other companies. In other words, it is a commercial version of the simulations that Nvidia was already using to test and also validate its own stand-alone software and hardware.
"There's no way we can drive and capture all the crazy things that happen on the roads."
Access to good simulation is crucial to developing autonomous vehicles, says Danny Shapiro, senior automotive director at Nvidia, in an interview with The Verge. "There is no way we can drive and capture all the crazy things that happen on the roads, there are trillions of miles that are driven, [but] many of those, most of those are very boring miles," he says. "After a certain point, you already master that."
This is when engineers have to study the so-called corner cases, or scenarios that do not happen so often. There are tons of these when it comes to driving, says Shapiro: cars with red lights, rage on the road, dangerous weather, intense sunlight at dawn or dusk. Do enough driving in the real world with test cars, and you will surely find these events and scenarios, but not often enough to learn how to handle them. For example, in the real world, you only have a few minutes each day to drive a particular road when the sun sets. In the simulation? "We can drive on all roads 24 hours a day at sunset and stage all kinds of potential dangers [other]," he says.
This is the reason why any company simulates autonomous miles in the first place. By lowering the entry barrier, however, Nvidia has made it easier for companies without the type of fleet size or financial backing that Tesla and Waymo boast of entering this space. In addition, Nvidia's business model as a stand-alone technology provider could help create a de facto industry standard for stand-alone simulation, if widely adopted.
Creating standards for autonomous simulation could be an important step for technology, because it is now difficult to assess the quality of simulations performed by private companies, according to Nidhi Kalra, senior information scientist at the non-profit organization RAND Corporation.
"The problem with any simulator is that it's a simplification of the real world," says Kalra. "Even if it stimulates the world with precision, if all it's simulating is a sunny day in Mountain View with no traffic, what's the value of doing a million miles in the same blind alley in Mountain View? I'm not saying that that is what nobody is doing, but without that information we can not know what it really means a billion miles. "
Kalra is a co-author of several studies for RAND on self-driving technology, including one in 2016 that sought to determine how many real miles would be needed to prove that autonomous cars are safer than humans. Kalra and co-author Susan M. Paddock concluded that driverless cars will need to drive "hundreds of millions of miles and sometimes hundreds of billions of miles" to make statistically reliable claims about safety. Because of this, they wrote, companies need to find other ways to demonstrate safety and reliability.
"When a company says we've driven so many miles in simulation," I think: "Well, I'm glad you have a simulator."
The simulations could serve that purpose, says Kalra, but there needs to be more context around those mileage claims. "If I say that I've played a billion miles of Grand Theft Auto, it does not make me a good driver," she says. "When a company says we've driven so many miles in simulation," I think: "Well, I'm glad you have a simulator."
Kalra says it's important to be skeptical about milestones driven by "simulated" miles that companies share unless they offer more details about what is being simulated. "Miles in the real world still really matter, that's where, literally, the rubber meets the road and there's no substitute for it," he says.
Photo: Sean O & # 39; Kane / The Verge
Knowing that Tesla and Waymo have accumulated the most miles in simulation and in the real world helps set the table for the discussion about who has the "most" data. But that knowledge is not enough on its own to really determine who has the biggest advantage. If Tesla breaks without self-driving without LIDAR, theoretically it could drive a software update for its customers that activates the switch.
But how will the company show that it is safe? Tesla has its own small fleet of test cars registered with the California DMV, but they drove zero miles in 2017. And throughout all the miles the company has accumulated with the current version of Autopilot on the highway through its Most customers have spent their time collecting data on the real-world application of semiautonomous technology technology that is once again being investigated by the National Transportation Safety Board after another driver died using the function.
Waymo could be in a better position to test safety through miles in the real world once it has a fleet of thousands of cars, but that could be difficult as it is still limited to a handful of locations. Even in the current lax regulatory environment for autonomous driving tests, progress in expanding these efforts will take time.
There is no perfect metric or definition of how "safe" these cars are
Another problem is how to define "security" to begin with. The only common metric that applies to all these companies is a so-called "disengagement", which records how many times safety drivers must regain control of the autonomous systems of a car. It is also an imperfect metric: it is only systematically cataloged by the California DMV, and it has been shown to be easy to fool because it has such a vague definition.
When the time comes for these companies to demonstrate to regulators or customers that they have developed a fully autonomous technology, the most likely measure that will be used to judge whether a company has developed a fully autonomous car is whether or not they are so safe or safer. than human driving. How to define that – the accident rate for X miles, injuries for X miles, or even deaths for X miles – is another matter.
As Kalra and Paddock point out in their study, this will be difficult to prove in real terms. But Kalra thinks that it can not be proven by simulation alone, at least not without a more complete and open understanding of the quality and rate of data that is collected. "We are probably going to see this technology deployed before having conclusive evidence about how safe it is," he says. "This is the problem". We can not show how safe the driverless cars are until we all decide to use them. "