There's Growing Evidence Tesla's Autopilot Handles Lane Dividers Poorly (arstechnica.com)

An anonymous reader writes: Within the past week, two Tesla crashes have been reported while Autopilot was engaged, and both involved a Tesla vehicle slamming into a highway divider. One of the crashes resulted in the death of Walter Huang, a Tesla customer with a Model X. The other crash resulted in minor injuries to the driver, thanks largely to a working highway safety barrier in front of the concrete divider. Ars Technica reports on the growing evidence that Tesla's Autopilot handles lane dividers poorly: "The September crash isn't the only evidence that has emerged that Tesla's Autopilot feature doesn't deal well with highway lane dividers. At least two people have uploaded videos to YouTube showing their Tesla vehicles steering toward concrete barriers. One driver grabbed the wheel to prevent a collision, while the other slammed on the brakes. Tesla argues that this issue doesn't necessarily mean that Autopilot is unsafe. 'Autopilot is intended for use only with a fully attentive driver,' a Tesla spokesperson told KGO-TV. Tesla argues that Autopilot can't prevent all accidents but that it makes accidents less likely. There's some data to back this up. A 2017 study by the National Highway Transportation Safety Administration (NHTSA) found that the rate of accidents dropped by 40 percent after the introduction of Autopilot. And Tesla argues that Autopilot-equipped Tesla cars have gone 320 million miles per fatality, much better than the 86 million miles for the average car. These figures don't necessarily settle the debate. That NHTSA figure doesn't break down the severity of crashes -- it's possible that Autopilot prevents relatively minor crashes but is less effective at preventing the most serious crashes. And as some Ars commenters have pointed out, luxury cars generally have fewer fatalities than the average vehicle. So it's possible that Tesla cars' low crash rates have more to do with its wealthy customer base than its Autopilot technology. What we can say, at a minimum, is that there's little evidence that Autopilot makes Tesla drivers less safe. And we can expect Tesla to steadily improve the car's capabilities over time."

Comments

Re:To be fair...

[Rei]

And the driver who destroyed said barrier was not on autopilot. Normal human error.

Check out the intersection on Google Maps and you can see what went wrong, both for the human, and for autopilot. The left line is quite distinct. The right line is rather worn. There is no visible crosshatching at all between them. Once a vehicle crosses the fading line, what looks like a "lane" forms around them, seemingly reinforcing that this is an acceptable place to drive. This happens only seconds before the barrier is hit, so there's not that much time to react to the situation. There are no overhead signs, just the road-level sign. In dense traffic, it's not visible until you're in the invalid "lane".

Any driver, paying attention, will of course not do this. But human drivers' attentions lapse, and that's a mistake that humans can - and recently did - make.

Concerning Autopilot, there's a big question as to what versions people are running. Walter Huang, at the very least, was almost certainly running the old AP. It's not clear what versions the YouTubers were running. There was a massive AP update that just started rolling out recently that makes a huge difference in quality. To the degree that I'm actually rather concerned about it. The more imperfect the system, the more attention you pay to it. I have worries that with the new system, it's gotten good enough that it's going to cause peoples' attention to lapse. Having to touch (with torque) the wheel at regular intervals helps, but I hope Tesla gets eye tracking in place soon.