How much has the recent interest and investment in self-driving cars impacted this estimate? 5 years ago, I would have said that level 4 automation is 20 years away. With all these billions pouring in, I'd cut that projection in less than half.

I'm certain that I won't be driving a car ever in 2030

I think it actually doesn't make that much difference. The problem, I believe, is that people are used to rapid advances in software. Lots of small iterations can take you a long way, but sometimes you just reach a hurdle which requires a breakthrough. A fully self-driving car would need a huge breakthrough in perception (which most people categorize under AI) in order to work.

Computer vision actually started out in the 60's because how hard could it be to classify objects in images? Well, decades later and it's still a hugely unsolved problem.

Robotics can be broken into 3 main categories: perception, planning, and controls. Perception is always the weak point. Robots just cannot see and understand their environments like people can. When you put a robot in a dynamic environment, they fail spectacularly. To give some idea of this, look at videos of robots falling down at the DARPA Robotics Challenge - and keep in mind those were made by teams of top roboticists with millions of dollars in a highly constrained environment (it's hard to say that obstacle course was "real").

I really, really doubt it's going to happen in a few years. I would love to be wrong though.

Last time I checked Google's self driving car was doing pretty well already without major breakthroughs. I suppose there still are problems when the weather conditions are less than ideal, but you can't compare it to /r/shittyrobots.

Many places have "less than ideal weather conditions" for 100+ days a year.

The car's performance on a relatively small stretch of road in California is hardly indicative of its potential performance on millions of miles of varied roadways, terrain and climate that make up the rest of the world.

How does it do in a snowstorm? Heavy rain?

well, hasn't google driven around actual city streets for many millions of miles? I could be wrong, but driving a car doesn't actually require breakthroughs in perception beyond current technology, The issue is that it just isn't economical for individual car ownership at this stage, but it could very well make economic sense for a fleet of taxis. Note, those taxis don't have to have the ability to go just anywhere, they could have predefined routes

> how hard could it be to classify objects in images? Well, decades later and it's still a hugely unsolved problem.

AFAIK Google's cars rely on radar to identify the position of objects, which I would think is vastly easier.

Of course, there's still the interesting challenge of classifying objects into traffic partipicants and obstacles.

>>sometimes you just reach a hurdle which requires a breakthrough.

When there are billions of $'s to be made. Generally there is way.

I understand from where you are coming. But if the trend grows, and one companies sees the billions coming. Others are forced to innovate. Also a whole new ecosystem is born.

See pre iPhone and post iPhone mobile industry. Sometimes you need one company to do something insane to drive others down the same path.

And interestingly, seeing challenges ahead, I don't think we will have 100% autonomous cars before 2050, if not later. 50 years ago, we were certain that perfect image recognition is just few months of work away, or at least enough money poured into it will solve the problem. It's 2016 and our best of the best image recognition algorithms can't even tell a zebra and a sofa in a zebra print apart. Since automatic cars are trying to solve the same problem I can't believe that in 2030 anyone will make 100% autonomous car, not a chance. Super advanced cruise control, like Tesla's autopilot, but better - sure, absolutely. But a car without a steering wheel? nope.

> It's 2016 and our best of the best image recognition algorithms can't even tell a zebra and a sofa in a zebra print apart.

State of the art offline CNNs have been doing classification quite well recently, arguably better than humans. See the Inception-v3 paper: http://arxiv.org/pdf/1512.00567v3.pdf

None of that changes the fact that they're easy to fool and have some pathological cases (that gambiting alludes to) that most humans consider quite dumb.

People like you are the best because it really adds to the enjoyment of great achievements when there are a healthy amount of naysayers to point out. Thank you, naysayer.

How many billions have been poured into speech recognition? I don't anyone here would trust controlling a car with Siri or Google Voice lest it misinterpret something you said.

Have you used Google Voice in the past ~year? I have a standard american accent so I might be the best case scenario for it but I haven't seen it miss a word I've said anytime I can remember in a year or two. I mostly use it in the car to reply and send texts, works even with the radio and ventilation on. I've used an Amazon Echo and that is right 99% of the time too.

Now, interpreting the words I say is a different matter, but as far as transcription I'd say we are at basically 100% right now and have been for a while.

> Now, interpreting the words I say is a different matter, but as far as transcription I'd say we are at basically 100% right now and have been for a while.

What? Sometimes I use Google's speech to text and a majority of the time it's a waste of my time. "Food near me" is not "fool's errand" (the irony isn't lost on me) and "gyms nearby" is not "Jim Dougherty", both are search terms I used last night.

shrug I don't have these issues at all. I just tried "food near me" and it worked perfectly fine. Like I said I might be the best case scenario since I speak in a basically standard american broadcast english accent.

Well, the thing is, English is not the only language in the world. Google's voice recognition in my native language, Polish, is absolutely abysmal.

I feel like that is more a product of the fact that the effort put into most other languages is greatly lower at this point relative to English, rather than English is particularly easy or other languages are particularly hard. It will all come in time.

Not effort - but data. A large amount of the dictation learning was done through users using and correcting the software. If you have 40,000,000 English users but only 80,000 Polish users you can immediately see why the software may not be up to par.

Google Voice to Text, from my experiences, is bloody amazing for Chinese. My Chinese is abysmal (think "never gets the tone right") and for short phrases/sentences I've never had Google fail on me. Maybe it fails on larger sentences but it would be difficult for me to believe.

Japanese is another language it does very well in. I can only count on one hand how many times it has made a mistake - and usually repeating myself a tad slower and with less slurring and it will pick it up just fine.

Well, my language (Danish) is pretty small and I would say Google voice recognition is notably worse than English. Sometimes it's way of pronouncing names is a bit weird but it is recognized when I said them.

So it might not work for everyone in the world but at some point soon it will properly be there for most national languages.

You'll be surprised to know how popular speech recognition is among younger people.

Amazon Echo is far from perfect, but it's the perfect example of the rapid improvements in technology + building an interface that "just works"

Same view here, current solutions are nothing more than Sunday drivers to me. They work well in good to near perfect conditions only where we actually need the least amount of self driving.

Wheel less driving, first show me it drives in the pouring rain at night when the roads aren't lit. Show it having the ability to obey posted speed limits, adjust to construction, and the like.

I am still surprised at how much current systems love to tail gate let alone don't suggest obeying posted limits.

Give me a car which self-drives for 4 months of the year, and requires me to drive it the other 8 months. That's still a massive advantage over the car I currently own.

Indeed but the context here are all the futuristic predictions starting with mass-unemployment of taxi and truck drivers (indeed that's what Uber is selling their market valuation on). If drivers are still needed 60% of the time, they're going nowhere.

The hype right now is so great that even if the reality is amazing it will be disappointing.

The disruptive effects still happen even if we don't get to 100%.

For example, a train of self driving trucks can be led by 1 in the front. Yes, I know, REAL trains exist, but they don't go everywhere. A train of self driving trucks could save a lot of money.

If self driving Taxis don't work in the rain, then Uber can just NOT drive them when it is raining! For example, Maybe the penetration rate will be very low in Seattle, but very high in Texas. Disrupting 20% of the Taxi market by eliminating drivers is still a very big deal.

Uber's main selling point has been availability when there are no taxis (hence surge pricing). If you can't get an Uber once it starts raining, someone else will eat their lunch.

And for trucks, will you let supermarkets go dry once it starts raining? Otherwise you need to keep drivers on retainer. And then you might as well just have them drive for minimum wage.

There would still be human driven taxis. The whole point of uber is that it is an on demand platform.

When it is sunny, the self driving cars are out and about, and the price is low.

When it is raining, the self driving cars stop taking passengers, and the price goes up, with surge pricing, which is handled by human drivers.

You don't keep drivers on retainer. You simply publish the price, and if the drivers like the price, then they can accept the job. If not, then they don't. If there is a shortage, then you use surge pricing until the shortage no longer exists at the given price point.

For trucks and supermarkets, if it is raining then you pay human drivers. If not, then self driving. BAM, labor costs have gone down by 50%.

They still have to work in the rain before gov't allows them on the road though..

>Indeed but the context here are all the futuristic predictions starting with mass-unemployment of taxi and truck drivers

Uber doesn't have any interest in self driving cars succeeding. It's a PR distraction to justify the exploitation of drivers ("you shouldn't complain because you are lucky to have a job at all when self driving cars are coming any day now").

Self driving cars that don't have drivers would completely destroy Uber's actual business model which is based on undercutting current taxi companies by shifting the capital, maintenance and insurance costs onto the drivers.

> If drivers are still needed 60% of the time, they're going nowhere.

Think about what you're saying here. If manual driving is still required 60% of the time there will still be mass-unemployment of taxi and truck drivers. These guys are not paid well enough that they can afford half the amount of income that they currently have.

So you fire all the drivers when summer starts. Then it starts raining. Who's gonna drive?

Most drivers earn money based on the amount of time they spend transporting people/things, correct? I don't believe there are many drivers who get a great base salary regardless of work output. Supply and demand would prevent people who tried to use driving as their primary income and the task would be taken care of by people willing to do it on the side when the demand is there

Interesting. A car that could "drone" in some contexts. Leave your parking lot and wait closer for instance. No need for high speed, complex collision avoidance. Only 10mph private neighborhood mobility. That could be a nice selling point. Something like the Tesla thing, just better.

NHTSA Level 3 ("Limited Self-Driving: Vehicles at this level of automation enable the driver to cede full control of all safety-critical functions under certain traffic or environmental conditions and in those conditions to rely heavily on the vehicle to monitor for changes in those conditions requiring transition back to driver control. The driver is expected to be available for occasional control, but with sufficiently comfortable transition time. The Google car is an example of limited self-driving automation.") is OK. That means that the self-driving system can recognize situations it can't handle (snow, potholes, crowds in streets) and come to a safe stop if the driver doesn't take over.

It's Level 2 ("adaptive cruise control in combination with lane centering") that's the problem. Therein lies the deadly valley - good enough that the driver can tune out, but not good enough to handle trouble. That's where Tesla is. Level 2 with strongly enforced hands-on-wheel seems to be OK, but don't let the driver take their hands off for long periods.

Both 2 years and 20 years are good estimates, depending on whether you're talking about a robocab that ferries people from their suburban homes to the train station when the weather is nice, or a vehicle that drive empty across the country in a snowstorm.

I think the bigger question is how you get straight to 4. Ford has resources, but I feel like they're betting against history here w/ regards to ML successes.

Practical, reliable ML stems from enormous data sets. And I can't see a way of obtaining the necessary data volume other than sitting in Level 3 for a product cycle...

I think the evolution will be using self driving features as a pure backup safety feature until they get good enough to go fulltime on their own.