Tesla Autopilot has driven over 1.2B estimated miles
hcai.mit.edu
hcai.mit.edu
I'd rather err on the side of underselling it by calling it ACC than overselling it by calling it Auto Pilot. But I'm not in marketing so what do I know?
The engineers that tested it undoubtedly turned it up to 11 and hauled ass around a track before dialing it back to a more "consumer friendly" setting.
As someone who heavily uses cruise control, and lives in an area with curves, it makes for unjoyed driving with patrons with adaptive cruise control, or non-cruisers.
I want to know how many miles are driven (both IRL and in simulation) per software release?
If you accept that the Autopilot software improves over time, then you accept that it changes over time. If it's changing, then there will be situations that will behave differently in the future. The implicit assumption is "the software will only make better driving decisions with each point release", but that doesn't follow! There's no proof that a situation that was driven safely by version 1.1 will also be driven safely by version 1.2 (unless it was tested in simulation).
If this means the pace of self-driving development would slow down, then so be it. I wouldn't let a driver turn on Autopilot with me in it.
At the moment, there remains some bugs to work out, as last night Autopilot decided to slam on the brakes at 150 km/h on the autobahn for no reason, but given how close it is to perfection on freeways, and how many improvements I see in the updates, I'll go so far as to say we should be requiring this technology in all cars like we now require emergency braking and backup cameras.
That's the problem with Tesla. They show autonomous driving, make big claims, talk about "autopilot".
And in the fine print they say in legalese that few people understand or even notice that this all isn't available right now and you must never leave your autopilot unattended.
Does everyone else pull over to change the CD or take a sip of water?
I may have worded my reply too strongly, see another comment about "second set of eyes".
But this is dangerous. When I open a bottle in my car, I know that it's a more dangerous situation than usual, I pay more attention, choose the right moment and see that I finish drinking quickly.
A Tesla driver can easily fool himself into "everything's fine, let me sip from the bottle for sevetal minutes and maybe search through my audio CDs".
I got the impression though that OP was stating that they used theirs as a "second pair of eyes" and added safety device when they do things in the car that require their hands and attention. This, to me, is perfectly fine. I know people who get around this in normal cars by driving with the knees, or just taking their eyes off the road for "just a second". Using autopilot has got to be safer than that.
I once had a wasp fly right into my face when I was driving, auto pilot would have been great, luckily I was on an empty road, but if I was unlucky enough to need to react fast as that happened I doubt I would have been able to.
You can't rely on the Autopilot to prevent accidents, it's simply not designed for this. So you have to pay the same amount of attention as if you were driving entirely manually. So if you're doing stuff that would be impossible while driving yourself, you're almost certainly not paying enough attention to the road and are at least negligent. Not sure about US laws, but usually there are laws against driving in a reckless or dangerous fashion, that can catch all kinds of behaviour.
You need to be always focused on the road and ready to take the wheel at any moment -- which in practice means you can't be using it at all, since "at any moment" means that 2-3 seconds can be all it takes between crashing and recovering from an autopilot error.
Holy shit! That's a hell of a caveat!
Indeed it is. To make it clear, it slams on the brakes bringing me down from 150 to 120 km/h, it doesn't come to a dead stop, so unless someone is seriously tailgating, then it's not creating any danger, it's just uncomfortable.
Remember, it’s the responsibility of the driver behind you to keep a safe following distance. If they rear end you, they were following too closely.
In this case we've established it was activated at random by the autopilot.
>Remember, it’s the responsibility of the driver behind you to keep a safe following distance. If they rear end you, they were following too closely.
That wont be on much comfort during a pile-up...
https://teslamotorsclub.com/tmc/threads/psa-careful-autopilo...
“But as I recall, the understanding then was that Tesla had a database of things like bridges, and the braking meant that the database was incomplete. So folks were told to file a bug report ASAP after the incident so that Tesla could fix the data.”
So! If this happens while on Autopilot, press the voice command button, and then say “Bug Report <speak bug report here>”
> Situation: When autopilot is engaged, it somehow sees some shadows/images, and it brakes pretty hard even though no car is in front of you. It does not come to complete stop, but it will go from 65-70 to 45 pretty quickly.
Conditions: Typically for me it has been driving in the dark (evening or early morning) and going under an underpass. I am not sure if others have seen this during the daytime or other conditions.
Another driver's description sounds like Google Maps getting confused by overpasses:
> It just happened to me today when I went under an overpass. The GPS based speed thought I was suddenly off the freeway and on the surface street overpass and said limited autosteer speed to 45. Luckily I had my foot on the accelerator so I could override it but it just jolt me a bit trying to slow down.
They’re pushing hard on the front facing radar to avoid accidents like the fatality of the Floridian driver who wasn’t paying attention when the tractor trailer crossed the road, and false positives are the side effect. Better to rapidly decelerate then be dead.
Even better to non-rapidly decelerate when it's not needed, since that can also have one killed.
Random as it concerns the actual road and driving -- in other words un-needed and unexpected for the driver (and other cars).
Why and how the Autopilot came to it, is really irrelevant, except to Tesla engineers.
If I crash because of Autopilot error it's no comfort knowing that it's doing its regular pattern matching just fine, it's just that something hasn't been seen enough to be registered as "as a known false positive"...
>So! If this happens while on Autopilot, press the voice command button, and then say “Bug Report <speak bug report here>”
Let me scrap the remains of my co-passengers and possibly my kids of the seats first. As for the bug, they'll learn all the details about it from my lawsuit.
Many automatic emergency braking implementations disable at high speeds, and/or ignore stationary objects, for this reason.
I get the impression that Tesla has made their automatic braking a bit more aggressive in recent updates, given recent publicity about cars hitting fixed objects while on autopilot.
No! That's a "game changer!"
Only if it works and has a better than human average driving.
Which might not even be enough itself.
If someone is a better than average driver themselves, then by using a "better than average" but lower than his capacity auto-pilot, they increase their risk, not decrease it.
Consider that the "averages" in accidents per miles also include drunks, elderly, people never really knowing how to drive well, people crashing in bad conditions (e.g. snow and heavy rain) and so on.
The average accident/miles for a fit good driver that usually drives in good roads in a state with good weather can be bigger than the average of the above.
So having an autopilot be merely better than the general average is not enough to make it "as safe".
>At the moment, there remains some bugs to work out, as last night Autopilot decided to slam on the brakes at 150 km/h on the autobahn for no reason, but given how close it is to perfection on freeways
Yeah, who cares if it can kill you at any moment with some bizarro sudden move, since it's "so close to perfection on freeways".
Thankfully, early adopters will help improve both the autopilots of the future and the human evolutionary pool.
This isn't a lowest common denominator issue, but rather addressing the swiss cheese model of driving safety. Autopilot catches problems that I, as a driver, might miss. At the same time, I catch problems that Autopilot might miss. The best example is emergency braking with this or any emergency braking system is truly superhuman. On the other hand, anticipating drivers that are about to veer into my lane is something I'll see before the car will.
Sure, if a great driver is not paying attention, then the safety of the car is completely left to the computer, but there's no evidence that this is done in large numbers. You can point me to some counterexamples on YouTube or that crash in Mountain View, but there's no evidence that this is done at scale.
> Yeah, who cares if it can kill you at any moment with some bizarro sudden move, since it's "so close to perfection on freeways".
I never said anything about almost killing me, and a blip of the brake pedal would only kill you in the event that someone is tailgating to an absurd degree, which is illegal and reckless by itself.
> Thankfully, early adopters will help improve ... the human evolutionary pool.
This is an unfounded statement. The death rate while using autopilot is still significantly lower than when not using it.
Too bad that, due to lack of awareness while delegating to the Autopilot, one can end up dead in the second case.
>I never said anything about almost killing me, and a blip of the brake pedal would only kill you in the event that someone is tailgating to an absurd degree, which is illegal and reckless by itself.
You didn't say, but you should have said about it. But you still see it under the very constrained situation where it happened, and the results it had. Not as an error that could happen elsewhere, and with more dire results.
As if the specifics of the road and the traffic at that moment where the only possible ones where this could happen.
While a fuller interpretation would be: "It has this BS error in judgement -- in what other cases could it be engaged and get me killed".
I'd expect people with no autopilot incidents to feel like it's perfectly OK. But to excuse it (because it's cool tech and they'll still ironing it out) when they have seen some, that's just reckless.
Like the unfortunate engineer who had seen the same issue that killed them, but brushed it off.
https://www.fastcompany.com/40551516/apple-engineer-who-died...
This is what happens when people enter a new market: They look for metrics, any metrics, and then optimize on those (because that is how business usually makes money in well understood fields) even if they are not really representative.
Autopilot seems to be around those numbers which is promising assuming continued improvements. Though it's not currently dramatically safer their is reasonable expectation it will improve over time.
Granted, I am approaching this from the perspective of a more relaxing driving experience not necessarily from a pure safety standpoint. People spend 20,000+ hours driving in a lifetime making that less stressful is a huge benefit even if they are still stuck in their cars and can't get work done.
That will be a huge adjustment that they seem to not be doing.
So any numbers of DUI deaths riding an 1990s semi-broken car is compared to an autopilot death in an expensive state of the art car with great engine, speed, and breaks.
How about we get the deaths per miles for non-coke users driving Tesla-like sports cars?
Even if misleading in the too cautious direction, that would be a better step towards safety than the numbers the car makers give.
Besides the same argument could be made for too much drug testing, against e.g. a snake-oil promoter.