Vorner's random stuff

Why we won’t have fully autonomous cars soon

As I’ve been browsing the Internet when researching EVs and cars in general (because I’ve wanted to buy one), I’ve came across a lot of speculations and hopes about future autonomous cars. There’s a lot of buzz coming from Tesla fans specifically, even though other car makers have some ambitions too.

I haven’t tracked each claim down to its origin, but these are something I’ve heard from various places, so let’s say these are the Internet „folklore“. Here are two picks:

They give the impression Teslas (with the Full Self Driving package) would get a level 5 autonomy. That is, one could send the car alone to pick up kids from school and expect it to get them back. The concept of a driver would cease to exist and the only reason a car would have a steering wheel (as an extra package) would be to have fun. Something like riding a horse today ‒ one doesn’t acquire that skill to get from point A to point B.

But I don’t think it’s going to be that way at all. Not at first, not soon. Tesla is on level 2 right now and expecting it to „jump“ 3 levels at once, or during a really short time when it’s somewhere on level 2 for 10 years is ambitious.

But I have further reason why I don’t believe Tesla or anyone would be getting level 5 any time now.

Note: I’m assuming specific large-scale infrastructure for autonomous cars isn’t going to be built. Not everywhere. Maybe some highways or so, but level 5 means it can drive in however broken, badly marked road a human can.

What I think the claims actually mean

Elon Musk is known to say audacious things and he’s good at marketing. He knows how to stir the waters. I believe these claims aren’t outright lies, but more of an Aes Sedai truth ‒ technically true, but somewhat misleading.

But even if I’m right, this would be a big step forward.

I think Tesla is aiming at level 4. That is, the car would be driving and navigating the traffic most of the time, could detect when there’s something it can’t handle and request assistance. It could pull over and wait for the assistance safely.

Such assistance would be more high level ‒ not fine adjusting of direction or speed. More like picking the right lane or correcting a speed limit. These can be done on the touch screen and don’t really need a steering wheel or pedals. They still need a driver, even though such driver doesn’t have to pay attention at all times.

Implementation of robo taxis

Given Teslas are connected to the Internet, such driver wouldn’t need to be physically present inside the car.

If enough people allow their unused cars to be used by such system, there would be always one close by, no matter how small a town or settlement. A driver wouldn’t have to be waiting inside each car, one would be assigned only once the car gets used. Some companies seem to be doing something similar as an experiment in the Bay Area. Tesla just wants to reuse the cars owned by folks.

By the way, I don’t have hopes this would be an open market. Tesla would, of course, be the sole operator of such service. Besides it being a good business, they could use such taxi driving as further training for the autopilot.

Why I’m sceptical to a fully autonomous cars arriving soon

I’m not going to talk about if Tesla’s approach of machine learning from examples of human drivers is better or worse than hardcoding rules, as seem to be the case of many other manufacturers. I’m simply going to argue that neither is going to give us level 5 soon because there are two more „cognitive jumps“ on the way, each one significantly more complex than the previous.

Technology may be speeding up exponentially, but so is the complexity of the tasks here. I expect one of the jumps to happen soon, but not both. Each one seems to take about 10 years at least, there’s no reason to think the most complex of them would happen in no time at all. It’ll be evolution, not a revolution.

So, let’s go over the steps. These are not based on the levels of self-driving, they are based on the complexity.

The past: Simple local systems

These are the assistants we have for longer time. They have some very simple, yet useful, local function. These are cruise controls, ABS, traction control…

These basically need only a simple feedback loop. With a cruise control, if the car is going too slow (measured simply by counting how many times the wheels turn per second), it adds more fuel / power. If it is going too fast, it uses less fuel / power. That’s it. Any coder after the first year of university is expected to be able to do this in a week. A geek with enough motivation will be able to do a mechanical version from Lego bricks.

They operate in well-defined conditions. They have few sensors wired to them, know what to expect and deal only with themselves.

The present: Making sense of the external world

One step further are things like adaptive cruise control (the one that slows down behind another car), keeping in lane, emergency braking or avoiding collisions.

For these to work, the car must build a model of the surroundings. It doesn’t need to particularly understand what these things around are. It’s enough to detect lines on the road, it is enough to detect that there’s something big moving and that if it doesn’t speed up by more than 10% in the next second, the car won’t crash into it.

That’s significantly more effort to make work than the above. Gathering sensor data from a lot of sensors, clearing it of noise, normalizing (compensating for different lighting conditions), putting all of that together and somehow tracking that the blob of pixels from the picture 20ms ago looks similar to this blob of pixels from the current picture. Then some more math, probabilities and physics in some 3D vector space.

It’s a lot of work, lot of tasks, but all of them are more or less well understood. It’s done research, it’s „only“ engineering. Put enough skilled people into the room with textbooks and computers and they’ll come up with the result in an expected time frame.

This is basically what gets us to Level 2 ‒ what we have now. An augmented driver. This is safer and more convenient, because a car:

Still, humans are needed for the higher understanding of the situation ‒ intentions of other drivers, traffic rules, etc.

The near future: Experience

In this step, the car needs to be able to step cognitively up from executing simple tasks (keeping in a lane while not crashing into anyone) to complex decisions. It needs „experience“ to be put inside it somehow. Going from „something behind me“ to „a police car with flashing blue lights behind me, which means I need to pull over and let it pass“.

It’s no longer just math and physics. This is getting „fuzzier“. Recognizing signs. Having statistics about the fact that a car flashing a direction light will at 95% chance go in that direction on the next intersection (the other 5% is forgotten light).

This is what the car makers are working on. In case of Tesla, it’s feeding a neural net with huge amount of human drivers doing stuff and trying to act the same. In other cases, it’s hardcoding rules. It means teaching the car about traffic rules, it means teaching it how to pick up the right lane on an intersection, it means teaching it that a lot of drivers don’t follow the traffic rules.

This is still work in progress and it is an active research. It’s on the level we have an idea how to go about that research and we have a pretty good idea that it’ll more or less work, with some unknown roadblocks on the way that’ll need solving. We know that we’ll make it work, we are not sure when.

It’s again vastly more complex stuff than the above. I mean, you can teach a tortoise to follow a drawn line. You’ll have really hard time teaching a tortoise or even a dog complex interactions on a roundabout or the right of way.

Once this works, it’ll allow for the car to do some amount of driving and ask for assistance once it doesn’t know how to deal with something. This gives us level 3 or 4 (going from 3 to 4 is simple from complexity point of view ‒ it only needs to pull over and park if the driver is not reacting and pulling over is definitely something it can do at level 3).

The level 3/4 definitions allows for it to work only at certain situation ‒ maybe just on a highway. But conceptually, this is only about „how much“ of this complexity is utilized. Driving in a city is harder, it needs more training than driving on a highway. But it is more of the same kind of training. More examples, more rules, bigger neural nets or bigger database of rules.

Given large enough training set, it can reach a state where a steering wheel is not needed, or the car can be operated remotely.

Well, given the right legal framework. But that’s a different story.

The far future: Context & improvisation

The above is for dealing with the usual. However, the reality contains a lot of unusual, broken, unexpected and surprising. For true autonomous driving, the car needs to learn to improvise and to understand context. Because it’ll hit a situation it has never seen before, was not present in the training set (or not prevalent enough to make it past the „noise“), is not described by the rule sets it has.

It isn’t enough to act „same as others“, based on „instinct“. It needs to understand the principles, the „why“.

Let me demonstrate. Imagine a red circle with the number 30 inside it. It’s a speed limit. Unless it isn’t:

Eventually, someone will make a sign that’s not standardized and expect it to be understood. Let’s say a crocodile inside a red triangle. For a human, that’s obviously a „Beware of crocodiles“ sign (That sign doesn’t exist in Europe, but if some crocodiles run away and start living in a lake, it makes sense. Other places, like Australia, might have their own version, but that one would look different ‒ maybe a yellow triangle). Because we know what crocodiles are, we know why they are dangerous, we deduce not to exit the car at that place. So, if we have a tire defect, we would probably go a bit further instead of stopping at that nice beach by the lake to fix it.

Sometimes, the rules are contradicting. There’s a dead-end one way street in Prague. Not kidding. Actual traffic-sign trap.

Some of these weird things will be prevalent enough to make an impact on the neural net, or someone will enter them into the rule set. But some of them are so rare they won’t. How many examples of someone having a defect near a crocodile area there are in the training set? None? Even if there was one, that’d get drowned into irrelevance by other things.

Humans are good at taking their whole knowledge into account, considering the current context and coming up with a solution (not necessarily a perfect one, but usually acceptable, by projecting the consequences and discarding obviously bad ones).

Neural nets are not capable of that (at least not today). Rule sets don’t work well for unknown or contradicting situations either. We actually don’t have a tool discovered for this yet. It’s not just about making a bigger neural net (unless the neural net somehow can encompass the equivalent of all common human knowledge). This needs another cognitive type.

Without this, we can’t have a true autonomous car. The human needs to be there somewhere, even if there’s a 0.01% chance the car wouldn’t cope. The human doesn’t need to be present in the car, and in fact, if the probability is good enough, Tesla or other automaker could provide the human operator as some kind of service behind the scene, part of the „Full Self Driving“ package (just hope it won’t happen at place without connection). That is, a car with very good level 4 capability and connectivity could be marketed and sold as a level 5 car.

This might be solved eventually. But this is of the kind of problem where we don’t even know how to start the research behind it. There’s always some time between researching something and making it production ready. For that reason, I don’t full autonomy in a year or five, not even from Tesla.