The driving part is usually normal algorithm code and not machine learning though. Machine learning is for object detection, what the software does with that knowledge is regular programming.

That's not quite right. There are learned models inside the planning systems of most of the sophisticated self-driving car companies. Here's a public video that gives a few high-level examples: https://vimeo.com/618478174

Full disclosure, I work for Aurora in the Planning group.

How do you know those models are safe? I know Google said they didn't use machine learning for the driving part.

Edit: To clarify, I wonder how you do this part:

> make safe, predictable decisions on the road.

ML models doesn't make predictable outputs. So you always need a system that can handle the ML model going haywire, since you have no way to prove that it wont.

This is probably a miscommunication issue. Waymo absolutely uses machine learning in the perception and prediction layers of the stack. Underneath that are a few layers of trajectory planning and controls code that don't necessarily need ML.

As for stack accuracy and predictability, they're active areas of research for everyone, and part of the secret sauce. There's lots of little mitigations scattered throughout everyone's stacks for various particular issues, but no silver bullet.

Very much so. The tens of thousands of hours of training is like the learning babies do while looking at jangling keys. Then the driving part is more static -- I'm not sure that it's entirely regular programming, there may be some training of models there. But my understanding is it's certainly not learning/training _while_ driving, and _almost_ certainly it's not even learning to drive. It's just doing more learning for object detection and tracking, like a baby watching the world from a pram.

So I feel like no matter how many autonomous hours these cars accumulate on the road, they have effectively 0 hours of "learning to drive".

I think the only (but very major) difference is that is that animals and humans learn general knowledge. Someone with 100 hours of driving also has 16+ years of additional general learning.

A driving AI has no general knowledge. It doesn’t actually understand what it is looking at.

It’s not possible to make up reasonable responses consistently when you don’t know what you are looking at most of the time, machine or human. Have you ever tried to solve a technical problem before you read the documentation covering the basic concepts? You’re basically throwing shit at the wall until something sticks.

> I'm not sure that it's entirely regular programming, there may be some training of models there

I doubt it, not anything serious at least. Machine learning isn't at the level where it can be used to navigate environments better than human coded algorithms can. Maybe if you had a datacenter worth of computing power to do it, like AlphaStar, but you don't have that in a car. And even if you had that I still doubt it, Alphastar had a perfect representation of the game and could simulate the effects of moves perfectly.

Agreed, I was trying to suggest that if it's not 100% static code, it's something like 1% models, 99% static code.

Well, to be fair, the human driver in comparison had to have a decade or two of learning before they could drive well.

I had about 16 years of walking and 14 years of cycling around gradually increasingly complex environments before I drove a car.

The car controls are not the challenge. It’s understanding what everyone else is doing and how to handle it safely while making progress and not surprising other people that is tough. And those skills are partially - only partially - transferable from your pre-car experience.

>I had about 16 years of walking and 14 years of cycling

There are young kids that drive carts and race and they don't confuse walls and solid objects. The reason we don't allow younger people to drive is because young people are bad at managing risk(I done some very stupid risky things with my bike as a teen, I could have broke my bones or neck),

Autopilot from Tesla that uses mostly or only cameras has the issue of identifying obstacles which is a much generic problem(that young animals and kids have it solved) so you don't need to appeal to the "age of driving" excuse for this guys.

Not to mention, it's not just driving that makes you a good driver. It's 16 years of life experience. You spend that time growing up riding around in cars and busses. Living in the world, knowing how things work, learning how people behave. Driving is so much more than just pointing your wheel the way you want to go and pressing the gas and avoiding objects.

So many people don't even know all the rules of the road, but still manage to drive just fine without accidents. There are different unknown rules of the road in all different places around the world, but most people figure out how to adapt to them in a very short time.

>So many people don't even know all the rules of the road, but still manage to drive just fine without accidents.

There are huge differences in insurance premiums that can't be completely unrelated to accidents. People like to talk about other factors, but still.

> I had about 16 years of walking and 14 years of cycling around gradually increasingly complex environments before I drove a car.

A lot of that has to do with humans not being born with fully developed brains. Animals can walk and navigate environments as soon as they are born, so there is no reason to believe that humans actually learns these things rather than those systems slowly maturing as we grow up.

The other part is that our pre-trained movement system is made for human bodies, not cars. So learning to drive a car is learning to use another mode of moving yourself. Also you have to learn traffic rules. But the other things like understanding environments etc you get for free for being a human, every single large animal can do the same. They wont learn the traffic laws or how to drive the car, but they know how to navigate environments without hitting things.

I feel like driving a vehicle is fundamentally different from "natural" activities, because it involves a large difference in how you deal with things ahead versus the sides.

Things in front are highly compressed, while things to the sides are not, and your ability to move sideways is reduced by mass and speed.

When I was a teenager, I went on a road trip where I drove like 12 or more hours in a day, which I could never do now. At the end, I got home and for a while I had a weird sort of tunnel vision that I don't know how to describe. It was almost like I was looking through a fish-eye lens or something; everything seemed distorted because of spending so much time concentrating on small lateral motions and things coming towards me at highway speeds.

Sometimes I walk routes that I also frequently drive and it reminds me how motor vehicles compress time and space.

Birds can fly through thick forests without colliding with trees, and flying also is very clunky and you can't change your direction quickly etc. If they collide they die, so the selective pressure is really strong for them to not hit anything. Cheetahs can also run at highway speeds on uneven ground and still do just fine.

I spent years riding a bicycle, but there are some fundamental differences between that and a motorcycle, never mind a car.

I'm doubtful that current ML techniques are that general.

> novice human driver with just ~100 hours of experience

So a 100 hour old infant?

Many animals can walk and navigate environments better at 100 hours old than todays self driving cars can.

Here is a 50 hour old horse, imagine if our self driving cars had this level of environmental awareness:

https://www.youtube.com/watch?v=NJEMh-_vaME

Horses (actually, all the equines) are able to stand within minutes of birth, and run with the herd within hours. This indicates how much is built in before any learning, a useful data point.

(Amusingly, while standing up is clearly built in - foals usually stand on the first try - lying down is not. I've seen a newborn foal try to lie down for the first time, and after moving a few legs, just collapsing to the ground. This makes sense - there's a survival advantage to being able to stand up quickly and reliably, while being able to lie down and rest quickly need not work as well.

>> This indicates how much is built in before any learning, a useful data point.

Animals have a heck of a lot of background knowledge like that. Statistical machine learning algorithms can't incorporate background knowledge [edit: not easily] and have to learn everything end-to-end. Hence why they remain dumb as bricks, compared to even simple animals (like insects, which are "simple" only compared to mammals, say).

This sounds like a start of a dystopian novel, where animal brains are being shoved into various robots. And that a researcher who's trying to stop it, ultimately fails and he's also turned into a robot.

I wonder what comes first, engineered biological brains that we can put in things like cars or that machine learning has progressed to the point we no longer need biological brains.

Biological brains are super cheap and efficient though. The brain of a fly can still do object recognition and navigate environments for basically free, weighs less than a milligram including cameras, and the factory to produce those brains is just putting a cell in a wet environment and it builds itself, the brain of an ant can build complex structures etc. It is possible metal brains will never beat those.

I seem to remember a short story by James Tiptree Jr. who had a somewhat similar theme. In the story it was children born with deformities who were given robotic bodies and then spaceships to navigate, if I remember correctly.

If only you didn’t have to pay them so much it could happen? (Only a little /s)

>> It's awe-inspiring how much more quickly a human brain can learn versus a computer.

Do you think we live in StarTrek future already? Human brain is the most complex thing in the entire universe, as far as we know so far. You compare it to a what, a computer?

When/if it happens, though, that we get actual level 4 / 5 asynchronous, fast locomotion. That process has to start somewhere.