After two years, a federal agency finally released the data it used to come to the conclusion that Tesla's self-driving software reduced crashes. It looks like the agency's statement may have been unfounded.
(TNS) — It was an extraordinary vote of confidence for autonomous driving by the nation’s top vehicle safety agency. Two years ago, the National Highway Traffic Safety Administration announced crash rates for Tesla cars dropped by almost 40 percent after installation of a self-drive technology called Autosteer.
“Forty percent. That was an eye-popper,” said R.A. Whitfield, director of Quality Control Systems and an expert in statistics. So “breathtaking and remarkable,” he said, that he didn’t quite believe it. “Extraordinary claims ought to be backed by extraordinary evidence.”
But when Whitfield requested the data, he encountered a thick bureaucratic wall at NHTSA, the taxpayer-funded agency primarily responsible for vehicle safety in the United States. On Nov. 27, 2018, after a federal lawsuit and almost two years, NHTSA finally released the data. Whitfield was shocked. In a detailed, 25-page report issued on Feb. 8, he said the NHTSA study violated basic principles of standard research methodology to the point where no conclusion of any kind could be justified.
Whitfield — whose 32-year-old Maryland-based firm, Quality Control Systems, helps companies reduce risk through analysis based on data and statistics — said the agency’s take on Tesla safety was “not well founded.” In scientific circles, that means bunk.
The agency did not dispute Whitfield’s assertions, but said it is “reviewing the report released by Quality Control Systems Corp. with interest and will provide comment as appropriate.”
The episode raises questions that go far beyond whether Tesla’s Autopilot is safe or not. It draws attention to the collection and transparency of data that will be crucial to crafting laws and regulations governing the use of vehicles that, in whole or in part, can drive themselves — and the extent to which driverless-technology companies can win public trust.
“The lesson here is a need for candor,” said Bryant Walker Smith, a driverless-vehicle law expert at the University of South Carolina. “What does it mean to be trustworthy in this field?”
Companies, regulatory agencies and politicians, Smith said, need to communicate clearly: “This is what we’re doing, this is why we think it’s safe, and this is why you should believe us.”
The technology behind driverless cars has advanced rapidly. In December, Waymo, the driverless division of Google’s Alphabet company, began offering a commercial robot car taxi service around Phoenix. Cruise Automation, owned by General Motors, is developing a driverless taxi service for San Francisco. Uber, after virtually abandoning its driverless program, has resumed research and development. With hundreds of billions at stake in a budding driverless car industry, every major auto company and tech company has skin in the game.
But the field won’t advance rapidly until the law catches up. Currently the federal government barely regulates driverless vehicles. That’s been left to the states, which are creating a patchwork of laws, with no nationwide consistency.
Congress tried to pass federal driverless legislation last year, but a key bill failed in the Senate.
“I think they failed in part because of a lack of trust,” Smith said. Recent polls show that Americans are leery about driverless car safety; more than half (52 percent) of those surveyed by Gallup last spring said they never want to use a self-driving car.
Lack of company and government transparency will only hurt, Smith said: “When you have a lack of trust in the technology, the companies developing the technology, and the agency regulating those technologies, government will be a lot less willing” to back driverless cars.
The issue of safety-data transparency gained widespread attention last March, after an Uber car in driverless mode — with an apparently distracted test driver at the wheel — mowed down and killed a woman walking her bicycle across an Arizona highway.
Most researchers and the general public have little data available to understand what happened in that crash. Uber’s reaction was to apologize, pay the victim’s family a legal settlement and avoid a public trial. The National Transportation Safety Board, or NTSB, released a preliminary report last May that noted that the Volvo car’s automatic brakes were turned off and that the woman was wearing dark clothes and was hard to see. But little underlying data about Uber’s driverless system was revealed.
Some states require limited data release, but its usefulness for safety studies can also be limited. For example, California requires companies that test driverless cars on public roads to provide annual “disengagement” reports to the public — a tally of the number of times a robot system had to turn control over to the human test driver, with a short explanation as to why. The California Department of Motor Vehicles issued 2018 numbers on Wednesday.
But the numbers, by themselves, don’t reveal much. A high number of disengagements might mean a company is pushing the edges of its system and making rapid progress — or that the system isn’t working well at all.
The latest controversy resurfaces an infamous Tesla crash in May 2016, when Tesla’s Autopilot – a so-called Level 2 system that legally requires constant human attention – drove a car under a semitruck crossing the highway, decapitating the driver. NHTSA investigated the accident. The driver was inattentive, the agency concluded in a January 2017 report, and the Autopilot technology was not to blame — even though the system could not tell the difference between the silvery side of the trailer truck and an overcast sky.
That NHTSA report included the analysis of Autosteer safety. Using Tesla’s own records, collected through its over-the-air software updating system, the agency looked at mileage driven in cars before and after Autosteer was installed. (Autosteer is the sub-system in Autopilot that can automatically steer the car into other lanes, pass other cars and turn onto exit ramps.)
Using air bag deployments as a proxy for crashes, NHTSA concluded that Tesla cars were involved in 40 percent fewer crashes after Autosteer was installed. If that result were applied to cars in general, thousands of deaths could be prevented.
Just after the report, Tesla Chief Executive Elon Musk tweeted, “The data show that the Tesla vehicle crash rate dropped by almost 40 percent after Autosteer installation.”
The company used the finding to support its contention that Autopilot makes cars safer. In March 2018, after a Tesla driver was killed when Autopilot drove the car into a concrete barrier, Tesla posted a response on its website that included a reference to the NHTSA study: “Over a year ago, our first iteration of Autopilot was found by the U.S. government to reduce crash rates by as much as 40 percent. Internal data confirms that recent updates to Autopilot have improved system reliability.”
Whitfield maintains that the data show nothing of the sort. The main problem he identified: NHTSA took air bag deployments before and after Autosteer installation to estimate the number of crashes per million miles. But most of the cars reported by Tesla were missing the miles the car traveled before Autosteer was installed. With no miles at all to add to the equation, but the same number of air bag deployments, any findings would inflate the crash rate for pre-Autosteer cars, he said.
For the small minority of cars for which mileage data were provided both before and after Autosteer was installed, Teslas were involved in 60 percent more crashes, Whitfield concluded. That could mean cars with Autosteer were more dangerous than cars without.
Whitfield emphasized that he has drawn no such conclusion. “We did not produce this data. We don’t vouch for it. We don’t know if it’s true or not,” he said.
But the exercise makes clear that NHTSA’s conclusions are “not well founded,” he said.
In his report, Whitfield said neither he nor his company has a financial stake in Tesla or any autonomous vehicle-related company or organization, and it was not paid by anyone.
His interest in the subject is motivated by public concern, he said: “Efforts to hide the crash record will impede progress in achieving whatever safety benefits advanced driver-assistance systems might ultimately bring.”
On Wednesday, Tesla issued a statement that read in part: “Our own vehicle safety data for Q3 and Q4, which includes data from roughly two billion miles driven in Tesla vehicles, shows that drivers using Autopilot were significantly less likely to be involved in an accident than those driving without using Autopilot.”
The company’s statement made only one reference to Whitfield’s paper, saying that “QCS’ analysis dismissed the data from all but 5,714 vehicles of the total 43,781 vehicles in the data set we provided to NHTSA back in 2016.” Whitfield, in turn, pointed out that he wasn’t dismissing data, he was pointing out that essential data were not provided.
Alain Kornhauser, who heads Princeton University’s autonomous vehicle engineering program, has another problem with the NHTSA finding: The data show that, if determination of safety is the goal, NHTSA is asking the wrong questions, he said. He notes that the NHTSA study didn’t assess whether Autosteer was turned on or off when the air bags were triggered.
“Isn’t the issue of safety of Autopilot the question of when Autopilot is engaged versus not engaged? The question is not whether Autopilot is available or not,” Kornhauser said. “Maybe we need more transparency. What we probably need is for NHTSA to release all of the data that they were given by Tesla.”
Tesla’s is hardly the only information NHTSA is keeping under wraps. Early last year, when General Motors petitioned NHTSA for a special exemption to deploy driverless cars in the U.S., law professor Smith filed a federal Freedom of Information Act request to see what GM was asking for. NHTSA has ignored the request, Smith said.
Last year, before the underlying data were released, the agency in a Bloomberg story called its own Autosteer conclusion “cursory” and said it was not assessing Autosteer’s effectiveness.
Congress will have a chance to revisit the data questions when it again tries to pass driverless legislation later this year. One of the groups that might be knee-deep in that effort — PAVE (Partners for Automated Vehicle Education) — has identified data transparency as a key issue in building public trust in driverless tech. The organization was formed last year by car companies, tech companies, safety advocates and disability rights groups to help educate the public. Tesla and Uber have not been invited to join the group.
©2019 Los Angeles Times. Distributed by Tribune Content Agency, LLC.