Autopilot not on during crash in PA, Musk says it would have prevented accident | Hacker News Reader

Autopilot not on during crash in PA, Musk says it would have prevented accident

electrek.co

21 pointsmaratd24 comments

I wonder if the driver thought the autopilot was on. A few times I've thought I'd enabled it, and only discovered that I hadn't, when I noticed the car starting to drift to the edge of the lane. It's especially easy to make this mistake when traffic-aware cruise control is on, providing some degree of automation but less than you think.

If you're using the autopilot to provide smoother control, and as a second set of eyes on the road, then thinking it's on when it's off isn't of much consequence. There's been speculation that the May fatality involved using autopilot as a replacement for real-time human oversight. Whether this driver had the autopilot on or off, the root cause – thinking that the presence of the autopilot makes it safe to ignore the road – was the same. If the autopilot was off and the driver thought it was on, though, then one of the suggested solutions – requiring hands on the wheel – might not have addressed this case. Or maybe any cases: I imagine someone who thinks the autopilot relieves them of responsibility is just going to rest a hand on the wheel while they zone out anyway.

> If you're using the autopilot to provide smoother control, and as a second set of eyes on the road, then thinking it's on when it's off isn't of much consequence

I understand the sentiment behind this statement, but I'm not sure I agree.

Compare with simple cruise control -- ordinarily, the driver needs to provide constant gas pedal input to overcome inertia at highway speeds. If for a moment the driver thinks that cruise control is on but it's actually off, the car will begin to appreciably coast, making it clear that it's off. The 'failure mode' of this scenario is rarely surprising or dangerous to the driver or others, and easy to recover from.

Whereas features like automatic lanekeeping, blind-spot warning, pre-collision emergency breaking... those are features that need to work unless explicitly disabled, because the failure modes all end in a potential accident.

There is more than one failure mode for cruise control.

EG: * Cruise control leaves throttle full open. * Cruise control fails to disengage. * Cruise control compression brakes on ice causing skidding. * Cruise control activates without driver's knowledge.

For an example see here: http://www.abc.net.au/news/2009-12-16/driver-recalls-freeway...

It's probably allowed in the T&C of owning a tesla, but is anyone else bothered by Musk calling out details of log files that "normals" probably aren't aware of?

Should people expect any privacy with regards this data? Maybe just share it with the NTSB and allow them to issue the statement that this accident was entirely not tesla's fault?

It's probably allowed in the T&C of owning a tesla, but is anyone else bothered by Musk calling out details of log files that "normals" probably aren't aware of?

It's quite common in investigations of plane crashes that the aircraft manufacturer is involved in the investigation and makes public statements (usually to the effect of "we've ruled out any problems with our plane... must be those darn pilots again"). The situation with cars is a bit different in that it's (usually) private individuals rather than professionals, but I'd say that the track records of car safety vs. aircraft safety definitely point towards following the model used for air crash investigations.

Data from black boxes about particular Toyota "unintended acceleration" cases became public, and some people were worried:

http://www.geek.com/mobile/toyota-acceleration-case-shows-wh...

In that case, Toyota's horrible programming practices do make it possible that the logfile doesn't fully reflect reality:

http://www.edn.com/design/automotive/4423428/Toyota-s-killer...

Elon musk is a stiff defender of his company and its practices. He will go on the offensive just to protect the name, nevermind politeness and formality.

At last some "positive news" with regards to autopilot. As much as I love the feature and it's progressiveness, I'm afraid the authorities might shut it down because of public safety concerns. Which might have some truth to it in the short term, though in the long run it almost certainly will be superior to the average human safety record.

On another note, it's interesting how often the autopilot is often blamed despite not being on. Thankfully that can easily be verified. The damage has been done nonetheless..

>> I'm afraid the authorities might shut it down because of public safety concerns

Yup. Dead people are certainly a public safety concern. Pesky dead people!

A premature optimization neglecting those killed in cars whose company CEOs didn't tweet about them.

Like the 30,000+ dead people from car accidents caused by human driver error every year?

Positive? Not really. So the driver turned it on and didn't complete the activation process. So why didn't the car detect that and slow down?

So did he try to activate it or not? If he attempted to activate it but did not complete the sequence of events the car should have been overt about warning about it. Perhaps before you can use it they should require people get trained on how to use it correctly?

The public should not be beta testing software of this type on public roads. There are many other methods to do so with better control, specifically trained employees in direct or indirect employment. It could be tested on fleet vehicles or the like.

Its still only driving in ideal weather and road conditions, while a nice feature its just an advanced cruise control. It certainly does not earn its name "autopilot".

Lets be honest, any other company doing this would have been accused of recklessness

Logs are a great thing, however I have had situations where once pointing out someones error (and obvious lie) that I was required to provide information to auditors how the data could not have been tampered with (difficult because lets face it, it could).

I wonder how its working in these situations, obviously a good/dodgy lawyer could easy spin the record as being tampered so I would want it rock solid.

That's why we have trials, right? If the facts were always perfectly accurate, there would be no work for juries.

In the real world... I imagine that the "dodgy lawyer" would provide a theory as to why the data is wrong, and Tesla's lawyer would provide a theory as to why it's right, and then it's up to the jury to decide what's true and not true.

Remember, a touch to the brake disables Tesla's "autopilot". Looking forward to the NTSB report.

Where is this data from? It would seem to require physical access to the vehicle. Was Tesla given access to a wrecked vehicle that's the subject of NHTSA and NTSB investigations?

Surely they must stream the logs back to Tesla as they're taken.

Things like "Car just had an accident" get sent immediately -- Tesla will call the owner and/or 911 -- but it appears that most logs accumulate on the car or are downloaded when the car is attached to WiFi.

When the car rolled over, the antenna was damaged and the data link lost, Tesla claims. So how did Tesla get the data?

I didn't lookup the exact Tweet, however based on the precise language of the linked article it is possible that the following occurred:

* Driver engaged the autopilot (assist)

* Driver did not have hands on the wheel long enough for the Tesla to disengage the autopilot.

* Driver may have instinctively reacted by applying acceleration to stay at speed, without further reacting.

* Vehicle crashed due to no steering control; auto-pilot not engaged and driver reckless for not being in control and situational awareness of the vehicle.

If I needed to get a new vehicle today (and there wasn't a waiting list) I think I'd probably get a Tesla; however I don't think I'd use the automation as I agree with the Google Self Driving Car stance on this form of technology. For that same reason I don't even like contemporary cruse control.

When the Tesla autopilot disengages because it doesn't detect your hands on the wheel, it takes a large number of seconds with increasingly obnoxious dashboard and audio warnings. It's hard to miss if you're paying any attention at all.

Is “contemporary cruise control” constant-speed, or traffic aware? Having driven tens of thousands of miles with each, I feel that the first increases the chance of an accident and the second decreases it (unless you let it lull you into ignoring the road). Tesla's cruise control has some blind spots (going around a curve into stopped traffic), but they're complementary to my own blind spots (the car in front suddenly decelerates; zoning out in stop-and-go traffic); my subjective impression is that the two of us together are much safer.

I wonder if we'll ever see Tesla's autopilot defaulting to a "watching and ready" mode, whereby it detects signs of inattentive drivers (hands not on steering wheel, drifting within lane, etc.) and turning itself on to bring the car to a safe stop.

That would be amazing for those who were having a sudden unexpected medical issue.

That's definitely the extreme case of "inattentive driver". There are different tradeoffs though -- if the driver is having a heart attack, you want to pull over and signal the emergency ASAP, whereas if the driver is merely distracted it's probably better to only take over when there's a danger to be averted.

Seems like he intended to slam on the brakes and instead slammed on the accelerator. And he turned the wheel left to avoid the obstacle.

If autopilot auto disengages, it is not the fault of the driver, but the fault of the design.

I wonder what we need more of ... better autopilot design, or better pilot training.

Reply on news.ycombinator.com