Self driving cars are still decades away

Almost every major automaker plans on having an autonomous car on the market with the next 3-5 years. I don’t buy it.

Despite Tesla’s brilliant marketing, lane keeping assists + adaptive cruise control does not make an autonomous car. Even if those two technologies can account for 98% of miles driven, the technology required to bridge the last 2% in order to create a vehicle that can go from any given point A to any given point B isn’t even on the horizon. Google’s self driving cars, as of a few years ago, reportedly can’t handle 99% of roads in the US, nor can they handle snow, heavy rain, any localized GPS outages, or unexpected and/or temporary traffic signals. The idea that these are small problems that are going to be solved in a few years is wishful thinking.

At best, we’re going to have long haul interstate truck routes that are completely autonomous, with ports on either end where a real person finishes the job. We’ll also see systems like Tesla’s Autopilot proliferate down to more affordable cars, but the law and the technology will still require the driver to remain alert and take over when the system can’t handle the situation, which will be at some point on every single trip the vehicle takes. This is the best we’re going to get for decades.

That’s what I think, at least, but clearly I’m a naysayer. Convince me I’m wrong :slight_smile:

I posted yesterday in a different thread, what might have been more appropriate for this one:

That’s a fantastic point. Any AI honed in the US would be worthless in a lot of countries, even ones without all the motorbikes, owing to a vastly different set of driving expectations. For that matter, AI honed in California would need to learn to adapt to Baltimore. Or, say, learn what a Pittsburgh left is.

Is this a sequel to last years Re “self driving cars” people talk as if this is a few years away. This seems nuts to me. thread?

ETA: I think it’ll be great to be able come back here in 2020 and see how many of the companies listed at the OP’s link made good on their plans.

I don’t know if it’ll actually accelerate (heh) the process but the second generation hardware on Teslas is gathering data on how actual drivers drive and comparing it to what it would do in a situation based what it is seeing from it’s sensors.

One good example is when a highway dips down to go under an overpass. There is often a large green highway sign on the overpass that looks like it is right in front of the roadway until the vehicle gets closer to the interchange and the road dips under the bridge. Human drivers don’t think twice about this, but Tesla’s sensors would freak out. So apparent it is learning this sort of thing as it gathers data.

I as well. I don’t remember the thread from last year but I see that you mentioned a company call Nutonomy. The 6 cars they had on the road in Singapore, nearest I can tell, were still test vehicles that had engineers behind the wheel at all times. One of them hit a bus. I can’t find much in the way of updates, other than their future plans.

But yes, it will be interesting to see how wrong I am. I expect a lot of automakers will roll out cars with limited capabilities like Tesla has now and declare victory, but we’ll see.

Ford has reportedly solved the problems of snow and total darkness; from the link in the OP:

I take it you missed this news: Rio Tinto makes Australia’s first unmanned heavy train haul with 100km Pilbara test run

Next year is only 11 weeks away.

I think it may well be 15 years or so before 75% or more of all vehicles on the road are autos, but I don’t think it’ll be more than 10 years before significant numbers of autos are on the road.

I don’t think it’ll even be 6 years, and I think most of them will be electric or hybrid engines.

Nor can I, but I did see that Nutonomy (an MIT spinoff, btw) got a $16M investment and they are backed by Bill Ford.

Well, what would you say was a situation where they could declare victory and you would agree with them?

Well, let’s take into account what the government expects from self driving cars vs. what we SHOULD expect from self driving cars.

Logically, self driving cars have to meet this bar: are they safer than human driven cars? if so, they are ready for the road.

The government and the public, since self driving cars are new and scary, basically wants it to be as safe as flying, which will probably be impossible for 100 years.

No what I think is absurd - the idea that someone could take picture of pretty much every house in the US, no the world, and all the miles of roads.

Wait - Google did that already.

This whole thing is moving much faster than I expected. Clearly the car companies feel they will be soon out of business without a good line of self driving cars. Sure the challenges seem daunting - anyone can write down a big list. But with hundreds of people working for a few years, they soon vanish.
Sure they won’t be 100% self driving. Sure there will be places they can’t go. But most people will drive back and forth to work everyday without touching the wheel.

I’d guess 25% of the market in 10 years, and I’m probably conservative.

The car companies will have spent billions, and I bet lots of people will be eager to get their hands on the cars. The government won’t stand in the way. In fact, they are being more supportive in allowing tests than I thought they’d be. This is something all sides can get together about.

Here’s whatthe data says. As of 2016, Google/Alphabet/Waymo (which is far ahead of the pack) is down to one error every 5000 miles.

This is a pretty low error rate, albeit not yet adequate for deployment. To the poster above who mentions how you can’t possibly take a licensed self driving car and put it in traffic elsewhere : well, maybe. I suspect in very crowded streets it might not do very well, especially if other road users are constantly cutting it off and taking advantage of an autonomous car’s inherent politeness.

However, the vehicle wouldn’t suddenly go on a killing rampage simply because it’s in a different place. It could be loaded with a different set of traffic laws (SDCs will have to load variant traffic laws anyways, as state to state laws are different), but the fundamental algorithms for obeying those laws, but prioritizing safety first, would remain the same.

Basically, none of the code would get changed, it would just have a different set of parameters for what to do at specific signals loaded, and a different dataset for actually recognizing local signage and signaling devices.

Are we going to see deployment in 2020? Well, every difficult engineering project has slippage, and this has to be one of the most difficult projects every attempted. I could see a few years of schedule slip, but they are not decades away.

When there’s no expectation of the driver having to do anything on a routine trip. That is to say, I should be able to plug in an address in my driveway and take a nap without any guilt should something happen while I’m sleeping, nor should I have more than a minimal risk of waking up halfway through the trip to find that my car encountered some unsolvable situation and has parked itself for the last 3 hours waiting for me to rescue it.

And right there in the introduction is my issue: “the vast majority on surface streets in the typical suburban city environment of Mountain View, CA and neighboring communities.”

They’re optimizing for a specific set of use cases around their headquarters. I’d be interested to see how they do when they plop one of their cars in Ames, IA for the first time.

There is a threshold that must be breached and once breached very quickly most new cars will be minimally at least mostly autonomous. (Sort of like the activation energy of a chemical reaction if you remember that graph.)

The threshold exists because the most difficult environment for autonomous vehicles is now when there are the fewest of them on the road. Human drivers are erratic, error prone, and hard to predict. Being able to deal with human craziness reliably is the threshold and obviously from there the technology only gets better while the challenge becomes less onerous as more of the other vehicles are autonomous (and predictable) and likely even non-autonomous new vehicles come equipped with vehicle to vehicle (V2V) communication that warns of its sudden changes and is warned of other vehicles’ sudden changes the exact instant they are implemented. Once those lines cross it goes fast for new cars sold.

My WAG is that “adaptive cruise control” (inclusive of stop and go traffic) and “enhanced safety features” with V2V as part of the package becomes fairly standard on new vehicles fairly quickly. Full autonomous being available within 5 years likely but taking off will lag until there is a critical mass of other vehicles on the road having those other features that make them less dangerous agents (and that provides the economy of scale for the basics of the technology). Until then vehicles capable of fully autonomous function will be a higher end option for individual owned passenger vehicles. The ride hailing market will OTOH adopt them more quickly as will the fleet trucking industry.

The whole idea behind machine learning and neural nets - the problem they’ve been trying to solve for 50 years - is about finding a general solution to a given problem.

Nobody wants the autonomous car to only be able to drive on streets that look exactly like company headquarters. And there are strategies for dealing with this, and google, at least, is up to hundreds of vehicles ranging a lot farther.

There’s also simulation training, and one of the more compelling methods of autonomous car training is what you could call ‘virtual backseat driving’. Basically, the computer in a Tesla not set for autonomy can be using the processing resources to observe what the human is doing, model what it expects the vehicle to do next and what it expects the driver to do next, and thus make gains in it’s abilities.

Have you ever driven in Mountain View and surrounding communities? I have. I’ve lived in Boston - they aren’t quite as bad, but not exactly quiet country lanes.

One problem for self-driving cars is that there will be a period of time, and it’s probably starting now, where the car is not quite good enough to get by without human intervention but it is good enough to encourage complacency in the driver. When the car does need help, the driver is not ready to give it. E.g. the Tesla crash. If the technology can progress quickly enough to get over that hump then I think it’ll be an easy ride from that point on and the way will be paved for genuine self-driving cars that can legally drive while the humans inside sleep or browse the internet.

I’d agree with this. The absolute minimum measure of success would be for my car to be able to take me home from work at 6am when my eyeballs are hanging out of my head and for me to feel comfortable succumbing to sleep rather than fighting it. There would have to be no requirements or warnings, either legal or from the manufacturer, for me to be holding the wheel or for there to be any expectation that I should have to take control.

Bonus points would be given for a car that can drop me at the office and then find a park, but I don’t expect that any time soon. I don’t expect it in my lifetime to be honest, but I would be ok with being proved wrong.

A train?

Road trains in the outback. An ideal use case for autonomous vehicles.

Marketing these cars will be tricky, since many people will assume that they can just get in and go wherever they want without doing any driving themselves. The elderly and disabled in particular will want them, since transportation can be so difficult for them to find. When there is an accident, and there will be, people being people, the lawsuits will be ugly because the driver will claim they didn’t understand the car wasn’t completely autonomous.

TV series about lawyers and doctors will start having plots about this soon, if they haven’t already.