[_red]

I have yet speak to any experienced embedded developer who actually thinks "full autonomy" is remotely possible in the near/mid term. The suggestions that it is possible, seems to come mainly from junior developers and snake-oil salesman.

We don't even have "full autonomous" trains, planes, or trucks yet.

[njarboe]

Would an auto-car that crashed at a rate 1/10th of humans be a successful "full autonomous" vehicle? You could save 30000 lives a year in the US by deploying such vehicles and probably more due to knock-on effects. But that would be about 10 auto-car deaths/per day. No way that works with the current media.

Unfortunately, it seems many people are not going to accept auto-cars with less than airplane like safety levels. Of course that will never happen. So yes, "full autonomous" vehicles are a long way off (probably forever), unless someone (Waymo?, Tesla?) can show they are much safer and some kind of national or state level laws are passed to restrict the legal liability of the makers and owners of such vechiles. Sort of like how ski resorts would not exist without special laws restricting liablity.

[fucking_tragedy]

> an auto-car that crashed at a rate 1/10th of humans

This is incredibly generous assumption, especially given the fact that cars with advertised Autopilot features, like Tesla's offerings, actually make accidents more likely[1][2]. Compared to the driver fatality rate in other luxury vehicles, Tesla's offerings nearly triple driver fatality[3].

This is like theorizing about a car that survives 99.9% of all impacts at any speed. It doesn't exist, nor is there any indication that it will exist, and it doesn't serve a purpose other than to prop up a contrived argument.

[1] https://arstechnica.com/cars/2019/02/in-2017-the-feds-said-t...

[2] https://driving.ca/tesla/model-s/auto-news/news/iihs-study-s...

[3] https://medium.com/@MidwesternHedgi/teslas-driver-fatality-r...

[njarboe]

No assumption, just a serious question. I agree no one is near that good yet. But it was a question that people should be thinking about. How good will be good enough to let the auto-vehicles operate on the roads?

[trickstra]

> This is like theorizing about a car that survives 99.9% of all impacts at any speed. It doesn't exist, nor is there any indication that it will exist

now THIS is a strawman

[trickstra]

quoting from your [1]

> So does that mean that Autosteer actually makes crashes 59 percent more likely? Probably not.

[hnhg]

This is a strawman. We're talking about replicating all of the abilities of a driver in all of the conditions they're supposed to operate in with approximately the same kind of performance. When we get to that point, then we can consider your question.

[Quick edit: a good question here though is what kind of drivers test is sufficient for us to even begin considering autonomous vehicles? Is that even answered?]

[trickstra]

why would that be a strawman? And why would we be talking about replicating the abilities of human driver? The ultimate point is to have safer transport, not copying a human driver with all of their quirks and flaws. Autonomous driving might look very different from human driving and still be safer. The way human drivers drive today is not necessarily optimal.

We don't want approximately the same kind of performance, we want much better performance, so it essentially must be different from the human driver

[Someone]

Level 5 isn’t only about safety, but also about universality.

At level 5, one expects a self-driving car to ride gravel roads, park in highly temporary parking spots, spot police officers and follow their orders, drive short distances on non-roads (e.g. to drive around a car pile-up), etc.

A car that recognizes those cases, stops, and tells it’s passenger “please help me out for a few meters” would be a fantastic accomplishment and very, very successful, but wouldn’t qualify as level 5 autonomous.

[sagarm]

I'm skeptical of autonomy in three years myself, but embedded engineers are hardly a convincing authority on the subject.

[trickstra]

Were you also skeptical about landing orbital rockets 5 years ago?

[kbenson]

I actually think landing an orbital rocket is probably easier by an order of magnitude, possibly even two.

Rocket science is a matter of applied physics, with materials science thrown in, and a bit of very well understood and straightforward software engineering.

Autonomous driving is a matter of getting AI models functioning to a well enough degree. My understanding of how this is done is you try to find more and better data to throw at it and tweak the models to hopefully make it learn better until the point it seems to pass your tests, which are whatever you've been able to come up with that you can think to test.

It's like trying to throw a pitch over home plate, but in one case you have a pitching machine you can aim and dial in the speed, and in the other you have a living pitcher. Only that living pitcher is an Orangutan you're trying to train.

One of those is a lot more art than science, and as such, getting well understood and reproducible outcomes that don't fall apart at a fundamental level when you add one more variable is harder.

[Someone]

Also, SpaceX’s landings are amazing, but looking at the numbers (https://en.wikipedia.org/wiki/List_of_Falcon_9_and_Falcon_He...): ”The rocket's first-stage boosters have been recovered in 35 of 42 landing attempts (83%).”

If Tesla’s cars were as good in parking as that, one in six attempts to park a car would lead to a fender-bender or worse.

Yes, that isn’t a valid comparison, but it does show that we accept way higher failure rates for rockets than we do for cars (aside: that also is the reason I don’t see space tourism become popular soon. If, say, the 20th or 30th millionaire who books a flight dies, the market will dry up rapidly)

[trickstra]

Seriously? Well then you just don't understand machine learning, robotics and the tyranny of the rocket equation. If you did, it would be easy to see that landing an orbital rocket is much harder than autonomous driving. The only problem is that nobody will pay 300M to get one fully self driving car after 13 years of research. And selfdriving needs tons of data which first need to be collected, which is what Tesla is doing at the highest rate than anyone in the industry by far.

Edit: one way to realize this is true is if you consider that the rocket already IS fully self driving. Everything after 1 minute mark before the liftoff is fully controlled by onboard computers, people are only there to push the big red autodestruct button if anything goes wrong.

Anyway, my point was something else. That something might seem impossible, and then after just 5 years it can be considered mundane. And I think that's what we will see with self driving too. Unless oil industry manages to manipulate public opinion in a way that stops Tesla before that.

[mhermher]

Wasnt that done in the 90s? Why be skeptical it could be done?

[derekp7]

You're thinking of the DC-X demonstration vehicle, which was a scaled down non-orbital vehicle. Similar to the early SpaceX grasshopper vehicle.

[trickstra]

I don't know, you tell me... "The SpaceX ORBCOMM-2 Mission successfully landed the first stage of a rocket during an orbital launch, a feat never before accomplished."

[gibybo]

"full autonomy" means lots of different things to different people.

Some people think of it as being able to drive anywhere at any time in any condition. This is unlikely to ever happen before AGI. What matters in Tesla's case is more likely being able to drive in some small number of places in specific circumstances without a physical driver in the vehicle (but possibly a remote one to handle unusual circumstances).

Now I don't think Tesla is going to even get to that in 3 years, but it wouldn't surprise me to see Waymo there.

[mhermher]

What part of "full" is ambiguous. Maybe your point stands if just "autonomy" was used generically. But "full" has a specific implication.

[gibybo]

I think we're in agreement, but Tesla has a lot of customers buying into the "Full Self Driving" promise. If you look at the "Full Self Driving" package currently available for $5k on their website, you'll notice that it doesn't even claim to be able to drive without a driver. It talks about doing what EAP can do today + doing it on city streets. What EAP can do today is basically adaptive cruise control with lane keeping and automatic lane change (which also requires you to hold the steering wheel).

[slg]

The primary cause of the ambiguity is that "Full Self Driving" is a marketing term that Tesla uses and not a unconditional promise. It is just like how cell phone companies promise "Unlimited Data" with a variety of caveats. So there is a question of whether people are talking about the primary definition of "full" or Tesla's marketing definition of "full".

[mandeepj]

> not a unconditional promise

Please try to avoid using double or more negates. It’s a bit hard to read.

[username223]

> a remote one to handle unusual circumstances

How could that ever work? The "unusual circumstances" that matter will usually require either action in less than a second, like an improvised detour sign, or situational awareness, like a human directing traffic around some random obstacle. At best, Google will be making cars that pull over to the shoulder and call for help in a few years.

[gibybo]

>At best, Google will be making cars that pull over to the shoulder and call for help in a few years.

Ya, that's pretty much what I was getting at.

[JustSomeNobody]

> ...some small number of places in specific circumstances..

Airport parking/shuttle service. They could even augment the AI by using markers in the roadways. That's about as autonomous as we're going to get in the next decade.

[r_smart]

Driving through airport terminals is possibly one of the most chaotic possible environments, with people almost suicidally throwing themselves in front of you, not to even speak of the cabs, double lane parking, security guards telling you to move and other things. It's probably one of the harder environments to automate for outside of inclement weather.

[kbenson]

Some airports have shuttles on private circuits to take you from one wing of an airport to another wing of the airport. The ones I've seen that have this either have a rail system, or a human driver. It would be much cheaper to just throw a bus with some software to track digital route markers and some visual matching for unexpected situations than to build a dedicated point to point system rail/track system (and maybe marginally cheaper than hiring someone, depending on other costs).

[paulcole]

> "full autonomy" means lots of different things to different people.

And don't doubt that Elon/Tesla use this to their fullest advantage.

[latch]

I generally agree that the lack of automation for seemingly simpler problems is good argument. However, there are a number of automated subway systems:

https://en.wikipedia.org/wiki/List_of_automated_urban_metro_...

[trickstra]

how do you call those subway lines without driver cabin? They are in service for a number of years in a couple of cities around the world