Slashdot Mirror
Uber Ordered To Take Its Self-Driving Cars Off Arizona Roads (nytimes.com)
After failing to meet an expectation that it would prioritize public safety as it tested its self-driving technology, Uber has been ordered to take its self-driving cars off Arizona roads (Warning: source may be paywalled; alternative source). "The incident that took place on March 18 is an unquestionable failure to comply with this expectation," Gov. Doug Ducey of Arizona wrote in a letter sent to Dara Khosrowshahi, Uber's chief executive. "Arizona must take action now." The New York Times reports: Uber had already suspended all testing of its cars in Arizona, San Francisco, Pittsburgh and Toronto. "We proactively suspended self-driving operations in all cities immediately following the tragic incident last week. We continue to help investigators in any way we can, and we'll keep a dialogue open with the governor's office to address any concerns they have," said Matt Kallman, an Uber spokesman. The rebuke from the governor is a reversal from what has been an open-arms policy by the state, heralding its lack of regulation as an asset to lure autonomous vehicle testing -- and tech jobs. Waymo, the self-driving car company spun out from Google, and General Motors-owned Cruise are also testing cars in the state. Mr. Ducey said he was troubled by a video released from the Tempe Police Department that seemed to show that neither the Uber safety driver nor the autonomous vehicle detected the presence of a pedestrian in the road in the moments before the crash.
SAE Level 3 automation should be illegal. Period. The backup driver simply cannot be expected to go from "no interaction with the vehicle for 1500 miles on average" to "rescuing the vehicle from an emergency".
Automation should be locked at Level 2 (ProPilot, Autopilot, Supercruise, etc, etc) until vehicles are at least ready for Level 4, if not Level 5. Level 2 = hands on the wheel, ideally with eye tracking, ideally making the user drive for themselves for at least a couple minutes every hour to stay alert.
And it should not be up to companies when their vehicles are ready to put them on the roads, as they're in a race to be seen as first movers. Governments should have their own review and testing processes, which involve both code audits (in the case of neural nets, audits of the net core and how the nets are trained) and real-world testing of simulated hazard scenarios.
"99 dead duelists of Dios on the wall. 99 dead duelists of Dios! Take one's ring, pass it around..."
According to the Arizona Department of Transportation, at least as of 2016 [azdot.gov], there were 952 fatalities in car accidents in Arizona, or approximately 2.61 deaths per day.
So far so good - now look up the number of cars in Arizona (about 2.4 million) and the number of Uber self-driving cars (200 across 4 cities). Now apply "appropriate precision" in Uber's favour - 200 out of 2 million = 1/10,000 of AZ cars are uber self-drivers. So, with 1000 fatalities/year, Uber get to kill someone every 10 years - they've used that up in one. (Of course, that's an unspeakably crude and dubious calculation, but its better than yours).
Then, of course, of those 1000 regular fatalities, many will be attributed to drunk-driving, speeding, texting (or other forms of reckless driving), non-roadworthy vehicles etc. all of which carry potential criminal penalties - including possible driving bans - so its not the case that nothing is being done about them.
Uber were allowed to test experimental vehicles on the condition that they'd have a safety driver ready to take over - and one thing that the video clearly shows was that the safety driver was not paying attention (to the surprise of absolutely nobody except, apparently, Uber). The video also shows that the pedestrian was crossing the road in clear line-of-sight, in a street-lit area, from left-to-right yet the car made no attempt to brake or swerve. If you believe that the video truly represents what the Mk 1 eyeball and/or the car's sensors could "see" then all that proves is that the car was going too fast for the conditions - outdriving its headlights - and the driver should have taken action to slow it down.
In a survey of 100 programmers, 111111 thought that duck-typing was a good idea.
It gets better, hereis what the street really looks like at night. About 1 second before the accident there is a yield to bikes sign due to all the bicycle traffic in the area.
Might be pointless but one recommendation I would have to the automated car companies is any warning sign, like yield to bikes, should have a visual and audible alert by from the car to "wake up" the human monitoring the car. Though given how often nothing would happen after that alert, it would probably get ignored after a while.
Shadow operations. Don’t control anything, just observe. If software action deviates from human action, review and analyize. It appears that this is the stage in product development that Uber needs to be in.
Once the shadow driving has advanced to a point where there are no negative deviations in a control environment, define that envelope and test rigorously within the envelope. Continue to shadow outside the envelope and slowly expand the envelope. Within test envelope, thoroughly validate performance with external telemetry— were there any cases where nothing bad happened, but the action taken should have been different.