First Tesla Autopilot death

[Marco]

I guess one of the advantages of working in tech is also understanding the limitations. Interesting moral question for engineers, is it better to provide "autopilot" mode and save hundred smart people from a collision even if one careless user dies? Or not allow anyone to use it until it is perfect?

For now you can simply use drive assist, where the car only partially intervenes in an emergency in a way which is still easily overcome by the active driver (ie. trying to steer away when you're cut off, resisting you when you try to cut someone else off). Making the driver do the work the rest of the time.

[zapta]

They both have low resolution compared to let's say 3D laser scanners. A laser scanner shouldn't have a problem distinguishing between a truck and an overpath.

Google use lasers in their car, but  lidar does not work in the rain.

In case of heavy rain there is plenty of time to safely transfer control to the driver. One of the jobs of the autopilot is to determine when it's unsafe.

[edy]

Florida Highway Patrol has published a sketch of the accident. Can't see any other possibility but driver distraction.

Looks to me like the Tesla (V02) was coming up to an intersection and the truck (V01) was making a left turn coming in the opposite direction. The Tesla did not "see" the truck because it was a similar color/contrast to the sky and decided to proceed straight through the intersection, going under the truck and chopping off it's roof and likely the driver's upper body was severely injured. The car then continued off into the ditch and smashed through some other things.

It does sound like a driver who wasn't looking ahead at the road.

A normal driver coming up on an intersection where you see people are turning left usually will slow down, be cautious and ready to apply breaks if the idiot turning left decides to cut you off. Also, the lights sometimes are turning yellow then red.... People see yellow and assume you will stop, so they start their left turn in your path. Sometimes people try to beat the red and accelerate straight into the path of a left-turning vehicle.

Either way, intersections demand WAY MORE attention than straight stretches of road. I can't see any fault of Tesla's either, except that they may have over-hyped the ability of these technologies. I am pretty sure there is a big warning screen that comes on when the car turns on that warns people though.

If anything good should come out of this tragedy, it should be that his death will cause more people to pay attention to the road and not trust their cars so much to handle things. We need a wake-up call and shake people out of this false sense of security. Even the naming of these systems has to change.

[apis]

As others have said, it sounds like these limited autopilots give people a false sense of security so that they stop paying attention. I think I would only be comfortable trusting a fully autonomous system (like the Google car) without any wireless networking that is connected to the autopilot computer (i.e. that can be hacked) and with a good track record for being safe (safer than a skilled human driver).

[G7PSK]

Not sure what the state of the art is now but a few years ago the RAF had trouble distinguishing between low flying helicopters and high flying trucks so if military grade equipment has problems how reliable will the sensors in cars be.

[Maxlor]

And a semitrailer is not going to cross the intersection very quickly. The Tesla driver could have seen the collision coming at least 5 seconds before it happened and then at least slow down or swerve if the autopilot missed the last point for coming to a full stop.

A statement by Tesla sayd they ignore high obstacles, otherwise the car would stop for low road signs hanging over the road. Could it be that the radar, mounted in the front of the car I believe, is angled so only horizontal obstacles are registered? It's how I could see it missing a trailer.

Are these car radars actually 2D, or just 1D strips? If the former, seems the problem would be fixable with some trigonometry in software. Otherwise, it'd be kinda hard to handle this situation, what with the camera apparently failing too due to lighting. Maybe with more situational awareness ("I just saw the tractor cross in front of me, chances are a trailer's following",) but things become really difficult there...

[helius]

This problem has been well-studied for decades, but the singularity zombies and trend hoppers don't study history, it's against their constitution.
http://spectrum.ieee.org/computing/software/automated-to-death

[continuo]

I would've thought the system would be restricted to lower speeds but it's apparently not. Maybe they will be forced to change that policy now.

[Towger]

I tend to see it the other way around: the less you need to concentrate on controlling the car the more you can concentrate on the other traffic.

The actuaries in our insurance companies disagree.  It cost more to insure an automatic car here, as they are statistically more lightly to have an accident.

[Maxlor]

The actuaries in our insurance companies disagree.  It cost more to insure an automatic car here, as they are statistically more lightly to have an accident.

I'm curious how they came up with that statistic, what with there being very little data on autopilot-controlled cars. Of course, it's an insurance's business to estimate risk even with little data to depend on, and the ones that get it right are the ones that grow. Still, maybe you can give some insight into the considerations?

[ez24]

Florida Highway Patrol has published a sketch of the accident. Can't see any other possibility but driver distraction.

.... going under the truck and chopping off it's roof and likely the driver's upper body was severely injured.

As in decapitated  - at least he did not suffer.  There should be a monument for him, for being the first.  As someone said, seems there should be a speed limit on this technology for at least a few years (maybe 50).

[dannyf]

It also has radar and sonar.

That's where tesla's explanation so far is fishy.

[nctnico]

Florida Highway Patrol has published a sketch of the accident. Can't see any other possibility but driver distraction.

Maybe but I'm wondering who caused the accident from a legal point of view. Over here (in the NL) you have to give way to vehicles which stay on the same road when taking a left turn. How is that in the US?

[Maxlor]

As someone said, seems there should be a speed limit on this technology for at least a few years (maybe 50).

I disagree, because high speed roads (highways) are actually rather safe (not much in the way of intersections, and everyone's going roughly the same speed), usually well maintained, and monotonous, perfect conditions for an autopilot. I don't think speed should be the criterium, it should be complexity. If you need to have a rule, make it something like, say, switch to manual control 10 seconds before intersections.

[zapta]

It also has radar and sonar.

That's where tesla's explanation so far is fishy.

Sonar is probably short range, low speed, for parking. For the radar, because of the low resolution, they selected a threshold to ignore overpasses, otherwise it would trigger correctly on that truck.

The public expects the machines to be perfect, not just same or better error rate than humans, and this can be expensive.


Edit: Tesla web site says 'enable after delivery for $3000' so this is basically a software option on top of the car's standard hardware. With this pricing model its hardware its underlying hardware and sensors are very price sensitive.

[bitseeker]

In a way, this is like the incidents with navigation systems where people blindly followed the verbal instructions onto railroad tracks, into lakes and ditches, through roadblocks, etc. Many were up in arms about how "dangerous" nav systems were. Unfortunately, the problem is the driver. Yes, the tech can exacerbate the problem in many ways, but the driver is still in command and responsible.

[bitseeker]

Maybe but I'm wondering who caused the accident from a legal point of view. Over here (in the NL) you have to give way to vehicles which stay on the same road when taking a left turn. How is that in the US?

Same (although rules can be different per state). The vehicle turning across traffic must yield unless they have a green left-arrow light that indicates they may turn (and a corresponding red light is illuminated to oncoming traffic).

[Rerouter]

I think this ad that just started appearing on Australian TV's is oddly relevant,



In it they do say the person that entered is at fault, but the other driver should still slowdown around intersections

[bitseeker]

Yes, quite relevant. Defensive driving is all about awareness and staying out of trouble, regardless who has the right of way or is at fault. It's better to be unharmed than legally correct.

[ez24]

As someone said, seems there should be a speed limit on this technology for at least a few years (maybe 50).

I disagree, because high speed roads (highways) are actually rather safe (not much in the way of intersections, and everyone's going roughly the same speed), usually well maintained, and monotonous, perfect conditions for an autopilot. I don't think speed should be the criterium, it should be complexity. If you need to have a rule, make it something like, say, switch to manual control 10 seconds before intersections.

I agree with you.  Make it manual when not on a freeway.  This accident was not on a freeway.  And it is on freeways that I wish I had something like this.

[Someone]

We know seat belts and airbags save lives.

Except there is still no academic consensus on this, so we don't know that it does save lives over a population without them. It will save some specific lives and is known to improve outcomes from identical crashes, but the overall effect is in debate. For instance increasing the safety of your vehicle has a negative safety outcome for all other people (not just other vehicles), but the relative rates of change are not accurately known and may vary with other factors.

[NiHaoMike]

For instance increasing the safety of your vehicle has a negative safety outcome for all other people (not just other vehicles), but the relative rates of change are not accurately known and may vary with other factors.

I don't see how safety improvements degrade the safety of others. Unless you're suggesting that with safety improvements, bad drivers are more likely to survive crashes and thus continue endangering others.

[chicken]

We know seat belts and airbags save lives.

Except there is still no academic consensus on this, so we don't know that it does save lives over a population without them.

Citation needed.

[ez24]

Except there is still no academic consensus on this, so we don't know that it does save lives over a population without them. It will save some specific lives and is known to improve outcomes from identical crashes, but the overall effect is in debate. For instance increasing the safety of your vehicle has a negative safety outcome for all other people (not just other vehicles), but the relative rates of change are not accurately known and may vary with other factors.

The reason is the world is flat.

[boz]

Lets face it driving a Tesla does not make you death-proof and however sad, more people are going to die in them in the future.

After the hype dies down (and there will be more as the lawyers haven't even got involved yet!) some good should come out of it. Tesla and other automakers will learn from it (dont trust human drivers) and the idiots who think auto-pilot is actually more than just a fancy cruise control will hopefully start paying attention to the road again.

I wont be refunding my model 3 deposit anytime soon as none of this changes the fact that the Tesla is still much safer and desirable than the average car.