Tesla "full" self driving beta is out, I talk about it

Topic: 
Tags: 

Videos how now emerged from the beta of Tesla's "full self driving" (really a city version of Autopilot.)

In this new article I outline various reactions to the limited amount we know about it today, what it means, and whether it's legal.

Read Tesla’s ‘Full Self-Driving’ Is 99.9% There, Just 1,000 Times Further To Go

Comments

Well stated case Brad. In viewing some of the FSD beta user videos I was impressed at how 'entertained' and amazed drivers and passengers were at the new algorithm. Having been a user of Tesla's autopilot/fsd system in my model S I'm looking forward to the more robust and eloquent system. I'm not expecting perfection or close to it for a number of years.

I hadn't given much thought to the permit issue. With Elon threatening to leave the state and continually dropping hints about how much better non-California workers are, my guess is they won't press the issue.

Remember when you said they weren't even working on a self-driving car?

They've done something truly amazing, and so far as I can tell they are far ahead of everyone else.

I've always said they wanted to build one. They aren't very far yet, and I and many other believe they will have to change their approach to get one. That view hasn't hoped.

They aren't very far yet?

They're finally at the starting gate. Everything done prior to this, using a 2D world view, was basically a joke. Musk promising Robotaxis using that code base was simply delusional.

I suppose you have software and hardware that can do what Tesla does, then. Where can I buy it?

Afaik, no one offers, at least not in the USA, anything even close to what Tesla offers. They are top of their class, not at the starting gate.

Tesla has built a car that drives itself, and they've released it worldwide. It's an amazing accomplishment.

What you call a joke is what enabled them to build it.

Never said Tesla isn't best at autopilot. They don't offer this new city autopilot yet, but may soon -- I've already bought it, though only because I got a special deal on it ($2,000) and I review such things professionally.

I can say all that and also say it's not at the same level as what Waymo has, and I believe Zoox and Cruise and some others have. You can't buy theirs and they don't ever plan to sell it to you though.

No, it's not at the same level as what Waymo has. It's far beyond it.

Waymo sells services in like 0.0001% of the world. And even there, the service is expensive and Waymo is losing money operating it. Tesla is available just about everywhere, for not much more than a regular car, and Tesla is making a profit selling it. There's no real comparison between the two.

At least in the USA, Tesla is the best at full self-driving. By far.

I rate it is far, far behind Waymo. I guess we will disagree. Tesla is "ahead" at the problem Waymo isn't trying to solve, and Waymo is ahead at the problem (Taxi) that it wants to solve and which Tesla would like, but is not that useful for a carmaker. It doesn't do Tesla much good to make a taxi that can only serve Phoenix or San Francisco because they sell to people who want to drive their car all over.

So they go at if different ways. But of course Waymo's plan is, once it establishes a working service in its early pilot cities, to then establish it in many more cities. Tesla hopes to let its owners hire out their cars as taxis, but their main business is to sell you a car for now. That needs a much harder ability to drive in lots more places because each customer wants to go somewhere unusual.

I'm not sure what problem you can say that Waymo is ahead of Tesla at. What they've built is useless so far, and there's no evidence that they're even a small fraction of the way toward building something useful. So far as I can tell, they're not even building a robocar. They're building an infrastructure-heavy remote-controlled level 3 vehicle.

Tesla, on the other hand, is now feature complete in building a full self-driving vehicle. It can be used as a taxi, or as a personally owned vehicle, anywhere, today. The hard work is done. Now all they need is to get the billions of miles of training, and that should happen in a matter of years for them. (It would take centuries for Waymo to get those miles, if they could stay in business long enough, which they can't.) At the end of the process, they'll have a level 5 robocar.

I doubt it would be years... Tesla can collect in one month what is takes most of self driving companies 10 years. If this is in shadow mode on other cars, is that is the plan this will be a lot quicker than years.

I think you'll be surprised by things coming from Waymo.

They have built a vehicle that you can trust to run vacant in suburbia. Nobody else has done that. It can now be a taxi. They say very soon you will be able to summon it like an Uber. Obviously they have just done it in one town. But soon they will have another, and another, and another.

They have built a vehicle that you can trust to run vacant (with remote supervision) in 0.0001% of suburbia. Just about anyone could do that. If no one else has (which is dubious), it's because it's worthless.

You say "soon they will have another, and another, and another," but their very methodology, relying on lidar and HD maps and remote operators and safety drivers during testing, makes that unlikely.

It's a lot more than .0001% of suburbia. And it took the most talented team I've ever worked with 10 years and billions to do it, so I think there is a case that not just anyone could do it.

Why does the use of lidar and maps in any way make it unlikely? That makes no sense to me. You act like maps are some intractable problem. Sure, they have to drive the service area in advance. Google has driven every single road in the USA several times by now with its maps cars (which 10 years ago had LIDAR added to them) and much of the world.

How much of suburbia is it?

I didn't say anyone could do it 10 years ago, but anyone could build a remote control car today. Yes, it's expensive. But it's not rocket science. If Tesla wanted to remotely monitor every car on the road, they could easily build something much better than Waymo that'd work anywhere with a good enough low latency network connection. (Maybe once Starlink is fully deployed they will.) If they stuck lidar in all their cars and had all their cars build HD maps all the time while they were driving, it'd be even easier. The problem is, it'd be too expensive. As you point out, Waymo has spent billions, and all they have is a service that charges more than Uber/Lyft and still loses money.

Lidar and HD maps are expensive. (Everyone uses maps, but Waymo relies on highly accurate, highly up-to-date, HD maps.) Every time Waymo wants to expand to a new city, they have millions upon millions in capital expenditures, then they have millions and millions more in operating expenses.

Tesla cars drive just about every single road in the USA every day. Waymo has a ton of catching up to do, and no business model for how to do it.

This is nothing like a remote control car, and does not suggest you could do this by remote control. But more to the point, a remote control car that needs constant remote driving is not much cheaper than just having a driver, possibly not cheaper at all. It can't scale.

Having remote operations centers is the correct plan. When you are in prototype phase, you have constant monitoring, because you are testing. Then, you upgrade it so you need one person for 2 vehicles, then one person for 3 and eventually one person for 100 and it's low cost and scaling just fine. Waymo has built a system which would do that, what you describe would not.

You say LIDAR is expensive but you know all the LIDAR companies are promising well under $1K and many are close already. You say mapping is expensive, just what is it you think it costs? And what does it cost per passenger-mile? I think you would have to be very bad it it have it cost more than a fraction of a penny per passenger mile in any scaled service, and who wouldn't pay that to give a higher degree of safety and performance and be on the road ahead of your competition?

It may not need constant monitoring (hence my comment that it's level 3), but it's still remotely controlled.

You say it's the right plan. I disagree. We'll see whose right, but right now Tesla seems to be winning, at least in 99.999% of the world.

Lidar is expensive. It's expensive today. The lidar that Waymo uses today is expensive today.

I don't know how much mapping costs. You tell me. But cost per mile in some hypothetical future is irrelevant.

It is not in any way "level 3." The remote operators are not able to take over. They can tell the vehicle to attempt to come to a safe state, and guide it on what lane to try and drive in, that's about it as far as I know.

Waymo builds its own LIDAR. They have not spoken about its cost in years. Do you know something we don't?

Why is cost per passenger irrelevant. It's the only cost that matters if you are a big player.

Apparently I know something you don't if you don't know that lidar is expensive.

Waymo is not anywhere near a big player.

I think your evaluation is fair of its .10 release and I think it's interesting they have been allowed to test it like this in this alpha build. But .11 is a huge improvement. 99.95% you could say. They use an algorithm to find their "Safe" drivers so I think they will be safe with this. I can see a time next month truly where you submit to this solution. If they can get this in the hands of 10,000 cars a small subset of their cars I can see a massive improvement very quickly, I just hope this initial coin flip by Elon doesn't come back to bite him.

I think expansion of this group will come down to number of corner cases per day. As the number goes down and the team is able to fix the corner cases quicker than they come in the group will expand, as long as there is a safe backstop... Tesla has many. If the worst case is an accident then Tesla might even foot the bill.

The corner cases become more rare and harder to find the more cars the more chance of finding it and if they can run this in shadow mode than it will happen even quicker

So what release are we seeing in the videos?

Tesla's data from drivers is going to be valuable to them, no doubt. But so valuable as they can leapfrog the others? Unknown?

You state that HD maps are safer, *even when they are wrong*. I disagree. Wrong data is dangerous, conflicting data makes the problem harder. Tesla maps are like the maps I have in my head: useful for routing, remembering that there is a turn lane or a stop sign, but in no means accurate, and therefore I use my eyes to confirm and to detail. Scaling is a problem that Waymo and others will have to confront.

Wrong data is fine as long as it is clear to detect when it is wrong. The more detailed your map is, the easier it is to notice when it differs from the world. If it differs from the world, then you are still better off than if you had no map because you can still use the map info which agrees with the world. The map not only tells you what you should see, it also tells you what everything you see looks like close up, and from every angle, and what it means.

I don't know about anyone else, but personally I use the "maps" in my head to confirm what I think I'm seeing with my eyes as well as vice-versa. (I also use them somewhat often to know about things I can't see with my eyes, like that there's a traffic light X feet in front of the truck that's blocking its view, or that there's a 35 mph speed limit sign blocked by a tree.)

Remembering what you've encountered in the past is incredibly important to driving. I'm not convinced that HD mapping is a good way of giving that information to robocars, though. It's really just a crutch to avoid having to program intelligence. It's a short-cut that might help you create an infrastructure-dependent faux-robocar in the short term, but does very little to help you build the AI you're going to need to build a robocar that can solve the general problem.

HD mapping is a software 1.0 solution to a software 2.0 problem.

Robots aren't at anywhere near the level of humans nor do we have any handle on when they might be. So the fact that humans don't use their memory in this fashion is not really apropos. But actually, human do use it, when we drive around blind corners, pull out of driveways and otherwise have things we can't perceive immediately with our sensors. But the real value of maps is helping the robot understand and well as perceive better. With the map, you are seeing everything static from many possible angles, and you get unlimited time to ponder it and to get advice on it from beings much smarter than yourself (if you're a robot.)

Now humans are good enough at understanding we can usually figure out what stuff is from a glance, and the roads are designed with this in mind. In fact, the road is full of signs to help us figure out what the road means ahead when road engineers have decided we can't figure it out immediately and need some help. We get a sign warning us about a sharp turn ahead, or a hidden driveway where trucks exit, or a good place to slow down or stop, or a bump or narrow road or 100 other things.

The map is the robot's version of that and at a higher level. Even a human, seeing a road feature as the come upon it, might not be able to figure out what it is the way you can if you get to look at it from all angles.

Coming upon the feature, you see only the front. Your map has decoded it from front and back and left and right, and much closer than you are now, and even with no human now has a better idea what it means. You don't just blindly follow the map. If you don't see the front as it is in the map, you then decide the world has changed, and the other views are also of no use, and you're back to where you were without a map. If you do see the front, it's pretty useful to have the map tell you everything about it now that you know it's there.

For robocars to replace the majority of human driving, they're going to have to be at and above the level of humans when it comes to the basic tasks of driving. Knowing which road features are important enough to remember and which road features aren't important enough to remember is part of that.

This is not some incredibly difficult thing to train a computer to be able to do. It was in 2010. But not in 2020.

The idea that you can take a shortcut and avoid that kind of understanding with things like HD maps, is misguided and naiive.

The point is they have to be above that human level. Humans can judge what features are important to remember, which makes it easier to match human levels. If the map can make you safer in just one situation, it makes you safer, and thus get to the goal faster. Of course, if it literally was just one situation it would not be worth the cost. But what about if it's a million situations?

As you've pointed out yourself, once you have a car that can drive itself without maps, it's easy to make maps. So, if an HD maps really would make things any safer (and I don't think it would), it's easy to make one once you've built a self-driving car.

What isn't easy is making a self-driving car after you've made a fake-self-driving car that relies on accurate HD maps.

If you can walk without a crutch, it's easy to learn to walk with a crutch.

It's not a million situations that are made safer by HD maps. Many humans drive perfectly fine without HD maps.

You benefit greatly from maps, of course. It helps tremendously to know where relevant things are even when you can't see them. But HD maps are only needed if your AI is crap. It doesn't help at all to know where every irrelevant detail of every irrelevant object is. It's just a cover-up for bad object recognition or other bad AI.

If HD maps help your car in millions of situations, your car shouldn't be controlling a vehicle. If HD maps help your car in millions of situations, your car is worse than the Uber one that killed someone.

Nobody has a car that can drive without a map, that's hypothetical. Better to say that as you work towards trying to build that, you are also building a car that can make better and better maps, which is to say, towards software that can construct a good enough map from one distant view of situation in real time on a single processor.

It's obvious to me that this means it is an easier problem to make software that can make that map with unlimited processor time, and with views of everything from every direction. A much easier problem. Much, much easier.

Humans use HD maps. They are called road signs which tell humans things the road engineers decided would not be obvious to the human brain. Humans are much smarter than robots and can do the rest on our own. Robots can't, they need more data than just the signs.

Perhaps nobody has a car that can drive without a map. I don't know. But surely the goal is to make a car that could drive without a map. After all, the car has to be able to operate when the map is wrong.

Yes, it's easier to make a map with unlimited processor time and views from every direction. So what? The car needs to be able to drive safely even when it doesn't have that. Like humans, it'll get it over time. As it drives roads more and more, it'll get more views, and there are a number of different ways it can spend virtually unlimited processor time analyzing those views in those few situations where that is necessary. Moreover, in the near term it's not just one car collecting this data, it's a whole fleet.

Maps are not bad. Maps are great. They shouldn't be strictly necessary for safe-enough operation, but they improve safety and even more so improve efficiency. You can safely go faster on a road you know better.

HD maps have no valid place in this, though.

Humans use HD maps.

I'm not sure what you mean by that. No they don't. Humans don't have centimeter-level precision memories of the world. Not even remotely close to that. Memories work completely opposite from the way HD maps work.

And there's a reason we've evolved that way. HD maps are stupid.

A car sort of has to drive when the map is wrong. But it's not as simple as that. A car will decide what to do when the map is wrong (knowing the map is wrong is fairly important, which being HD helps with a lot.)

It can:

  1. Decide to drive slightly less safely in the zone where the map is wrong
  2. Decide to pull over and get remote assist, or turn around where the map is wrong
  3. Decide in advance to avoid roads on reports indicating the map is or might be sufficiently wrong to require #2.

The key to #1 is this. The map makes you a bit safer, not a lot safer. But a bit is a lot when multiplied over all driving. Let's say you are 98% as safe (whoever you want to quantify that) without the map as you are with it. The car with the map drives at the 100% level 99.9% of the time, and drives at 95% safe 0.1% of the time when the map is wrong. (And perhaps less than 0.1% using the other 2 approaches.) That's a reasonable total safety record.

Drive without a map and you're at 98% all the time. You have taken a serious safety hit.

But the reality is, unexpected deviations from your map will be extremely rare. Because that only happens when you are the very first car to encounter the change. Changes are moderately common, but being the first person to see them is very rare. After that, it's not a surprise any more.

Humans do use a major HD map -- it's called the lines on the road. This is a map that is conveniently drawn right on the territory. Robots don't need the lines. If a robot can find the edges of the road and knows how many lanes there are, a robot could come up with very precise models of where the lanes are without painting them there. Humans can handle it without the lines, but the lines make us a little bit safer, let us go a little bit faster. Robots do not need them, other than to keep the humans in line. Robots just need to agree how many lanes there are.

The lane markings, the signs that warn you to slow down or that the road curves ahead, these are the maps for humans. Robots need more detail in some areas than we do, less in others.

Drive without a map and you're at 98% all the time.

Absolutely no one is suggesting driving without a map all the time.

Humans do use a major HD map -- it's called the lines on the road.

That's not a map. If the lines on the road are an HD map, then every robocar has an HD map. But no, that's not a map.

Robots don't need the lines. If a robot can find the edges of the road and knows how many lanes there are, a robot could come up with very precise models of where the lanes are without painting them there.

In a world where all the cars on the roads are robocars, that'd be fine. That's not what is being built, though. We're building cars that can drive alongside humans. What is important for that task is where the lines actually are, not where they hypothetically should be. (If the lines are drastically wrong, perhaps the flow of traffic will ignore them, but even then it's important to know where the lines are so you can anticipate the confusion from other drivers that is much more likely.)

The lane markings, the signs that warn you to slow down or that the road curves ahead, these are the maps for humans.

They're not maps, and they're not just for humans.

Yes, people are suggesting driving without a detailed map almost all the time. The Tesla approach has only a navigation level map most of the time, with occasional extra detail at certain points where they know their vehicle might get confused.

But in general when people say "Drive without a map" they mean drive without a map, almost all the time if not all the time. If you take it to mean "Drive with a map anywhere the road is complex and a more limited map when it's simple" I call that a map-based approach. Yes, you don't need a detailed map for a long stretch of interstate in any approach.

The lines on the road are the human's analog of a robot's HD map. The road engineers calculate them because humans can't easily do it, and then humans read it as they drive. Ditto the signs that guide you as to what speed to take and where to slow down. Cars are welcome to read them, of course, and many do.

A map is "information stored about the road and what you need to know about it to handle things you can't do with the same reliability from pure perception." Yes, it is an analogy in that the lines and signs are picked up using perception (in some cases) but they are an overlay layer on the road, not the road itself.

I know we need lines for humans today. I am just saying that humans and robots both like to have extra information about the road calculated for them in advance, such as where the lanes should be. That's what maps are -- a pre-calculation of information you want about the road.

When I was there Waymo's cars did not use the lines on the road, and I think this is true of many cars today. They are only there for the humans. Now, the human mapmakers used the lines on the road in deciding where to record the lanes as being, but the vehicle did not use them. By now, I am sure there is code to read the lines directly which you want when you're off map, like in a construction zone. If you're on-map, you are better to trust what the human map-maker confirmed are the boundaries of the lines than to use your more limited perception abilities to figure it out.

Absolutely no one is suggesting driving without a map all the time.

Tesla uses maps. They map numerous road features. They try to map everything that is important to map. Obviously they haven't mapped everything important to map in the entire world (no one ever will), but they map a lot. As you drive around, the car remembers important things that it sees.

The lines on the road are not a map! A map of the lines of the road is a map. An accurate-to-the-centimeter map of the lines of the road is an HD map. It's also an unnecessary, fairly useless map.

When you were there was many years ago, right? If Waymo still isn't looking at the lines on the road, then they're much worse off than I thought. The lines on the road define where the lanes are. They define what types of lanes the lanes are. They define what type of lane changes are allowed. If you're saying that Waymo cars use maps generated from the lines on the roads, and not the lines on the road directly, then (1) that's unsafe, as the lines might change, and (2) they're still using the lines on the road, albeit indirectly.

Any "robocar" that relies on humans to make maps for it (and can't drive in places that aren't mapped) is not a robocar.

Are for the same purpose as a computer's map. That's why I call them the human equivalent of the HD map. They show you precisely where the lane is, as calculated by the road construction team. The lanes don't physically exist.

I am sure Waymo is likely to be looking at the lines on the road. If the lines have changed, you need to understand the new ones. The map makes sense because not only does it confirm where the line is (even after the line has worn out) but it tells you the meaning of the line, and it does it better than you can do looking at it through a camera in the distance, because it has data from other passes over the same road, unlimited CPU and neural network size, and the ability to get human QA on that meaning.

The lines on the road are traffic control devices. They define where it is legal to travel. They're like stop signs, or red lights, or speed limit signs. Sure, you can map them, but the law is the physical reality, not the map.

I'm also sure Waymo is looking at the lines on the road. They're trying to build a robocar, and they're not completely incompetent.

Add new comment