Self driving cars are still decades away

In My Humble Opinion
963

I’m not sure everyone is aware (though I’m pretty sure Sam is) of the Gartner Hype Cycle that is updated each year for a variety of technologies. It’s a nice way of looking at the utility of technology on the basis of marketplace acceptance (instead of just technology maturity). Here’s the 2021 AI hype cycle diagram. Note the relative positions of Autonomous Vehicles and AGI.

It’s important to distinguish between AI and Autonomy. In simplistic language, AI is technology, autonomy is capability (that uses AI and other technologies).

The most succinct criticism of AI for Autonomy was when a professor speaking at one of our workshops stated (paraphrased from memory) “AI is great with images, but bad at scenes”

964

I can’t say I’m impressed. What’s their excuse for having zero items in the “Plateau of Productivity”? My supposition is that they are going with the old AI trope: “as soon as AI solves a problem, we cease to call it an AI problem”. Chess computers ran into that first (“they aren’t actually intelligent; they’re just searching a database”), but the same principle has been applied to everything.

Their placement in other areas is also dubious. “Synthetic Data” in the innovation trigger category? That makes zero sense; it should be in the “plateau” part. Everyone using deep learning for anything is already depending heavily on synthetic data. There’s continuing research of course, but it’s well-established technology. Or “deep learning” straddling the line between the “peak” and the “trough”? Again, nonsense–deep learning in its many forms is not only in heavy use in production, but continues to get better at a rapid rate.

Yet another “ok, so AI can do that, but it requires real intelligence to do this other thing.” Repeat ad nauseam for each new thing that AI solves. It wasn’t long ago that AI had no understanding of images at all, and that questions like “does this image contain a cat” were not only unsolvable, but at a point where researchers couldn’t even conceive of what a solution would look like. Today, I can search on my phone for my cat’s name and it will show me a list of every photo I’ve taken with the cat.

965

So no Positronic robot brain? No HAL? Thankfully! No Terminator?

I seem to have wasted my time and money reading all those SciFi books in my youth. Here I am decades in the future and there is no Teleporter, no holidays in space. No hologram room.

Not even a car that can drive me home when I am drunk.

I am unimpressed! How does this trough of disillusionment last?

A few more decades?

966

I can only repeat that this graph is not a measure (directly at least) of maturity or utility. Here’s a link to Gartner’s description of the Hype Cycle:

Gartner Hype Cycle Research Methodology | Gartner

And I’ll note that the graph also has an estimate as to how soon the technology will reach the plateau of productivity. Synthetic data is 2-5 years away, while Autonomous Vehicles are 10 years away which certainly doesn’t support he linearity in the horizontal axis that one might automatically think is there. Not counting startups, research, development, and prototyping, has synthetic data reached “mainstream adoption” and “…broad market applicability and relevance…” in the commercial marketplace?

The paraphrased quote was from about 6 years ago and was from a well-known researcher in AI and Autonomy. For me, it crystallized a problem that I don’t think has been solved yet. Humans process scenes (and create scenarios) when performing tasks like driving, and can do this using remarkably little training data (often too little). Deep Learning struggles to turn images into a scene (and projected scenarios) without significant training data and processing.

I think this limitation will eventually be overcome, but my guess is that it will take a breakthrough beyond the availability of more capable processing (which is really how I think image recognition with AI got off the ground- evolutionary improvement in algorithms and a huge increase in processing power and storage capability to run Deep Learning since the emergence of Deep Learning theory and algorithms in the 80’s). I think there are still architectures and algorithms that need to be invented, though I’ve underestimated how powerful processing chips could become in the past, so don’t take my guesstimates as gold.

967

Yes. Everyone using deep learning for autonomous driving (whether they’re targeting L2 or L5) has built sophisticated driving simulation systems. Basically no different from a high quality driving game, but with an emphasis on absolute realism. This data is used to train shipping products.

There is the danger of course that the net will fit itself to some peculiarities of the simulation, which is why real data is also important. But massive amounts of the training happen on synthetic data, and each new “interesting” incident that comes through the real datasets sets added to the synthetic one–but with the addition of happening with varying lighting, weather, etc. conditions.

Everyone is also using data amplification, which is arguably a subset of synthetic data. Take some real data, but transform it in various ways–mirror it, rotate it, change the color balance, add noise, etc. It can only go so far, but when collecting data (or just labeling) is expensive it can make your dataset go much farther.

Also in the category are adversarial networks. Essentially, you train two AIs: one to consume the data, and the other that tries to produce data that “fools” the first one. Photoshop has been shipping this for a couple of years now for their various “neural” filters. Probably you’ve seen a zillion photographs already that have been altered (to remove unwanted portions, or to extend beyond the borders), etc. using these features.

7 years ago, Go was in the category of “yes, computers have solved chess, but Go requires special human intelligence that we are many years from emulating”. 5 years ago, AlphaGo completely dominated humans.

At any rate, I agree that humans are currently much better at learning from limited data than AI systems, but this is improving all the time. And, well, it may not matter. Enormous datasets are available. Arguably, the human ability to extrapolate from limited data may prove a negative, as we can see by various forms of faulty generalization. Humans don’t have a choice because our possible dataset is very limited. An AI can be more careful about generalization if it has far more data to work with.

No doubt, but neural nets still have a lot of life left, and it would not shock me if they were never really supplanted. New topologies, new training systems, etc., of course. But still with a simple weighted-sum-and-nonlinear-transform at the core (just as computers still have a simple transistor at their heart).

968

Here’s a blurb on what Tesla is doing when it comes to simulated data (from 2021):

Although clearly this is primarily intended for self-driving, the tools they’re using also apply to basic Autopilot.

969

I remember AI or something like it being taught in the 1980s. There was much excitement about Expert Systems. But there was a distinct lack of storage and processor horsepower back then. I read recently that there have been at least four times in the past when it has been the exciting new technology of which great things are expected. It still struggles to interpret patterns in data. I am not much impressed by voice recognition. It is still very rudimentary despite the powerful computing resources we can reach on the Internet. How many people are dictating their responses on this here forum? I’ve used language translators and they are better than nothing, but they are not very accurate and make embarrassing mistakes. I have a friend who is a translator who says they can be a productivity aid, but it requires a human to ensure the meaning is accurately passed on. This is very important in professional translation, especially fields like health and law.

I am sure great strides have been taken in bounded problems like games. So they beat humans at Chess and Go. But real world applications are seldom so conveniently bounded. I can see how maybe moving little trucks around an automated warehouse might work. Though even those places require human pickers. But a city road network? They were designed for people. The trains aren’t even fully automated and that is a much more bounded system.

We are still learning to handle the deluge of data from sensors within a car. Making sense of the roadways with signs designed for humans is a different ballgame. It might happen in a clearly marked parking lot or garage or depot, for specific models of commercial vehicle within a decade. But out there on the wild highways full of crazy human drivers and ill maintained roadways and signage and weather? That is far too ambitious and the wrong place to start. Better to solve the easy problems first.

I am not even sure we are going to be actually driving in some cities in the future. The trend is to tax cars and parking to reduce the burden on city roads. Many are running at capacity. Drivers may be priced off the roads. Maybe confined to the suburbs and out of town rather than urban gridlock. Does this question arise from the frustration of an urban commute? The desire to relax and let the car do the driving as you spend too many hours sitting in a traffic jam? A more comfortable traffic jam experience does not sound like progress. The problem is not the car but the infrastructure it uses: the roads and the parking. These are not free, someone has to pay.

I suggest that the technology to find faster and quicker ways to tax car use will arrive long before self driving. Eventually we may see something like a taxi meter in each car to inform the driver of road pricing, zone entry charges and parking fees that are going to be automatically deducted from their bank account. It will probably be an app on your smartphone. The spin off from Identifying vehicles and their location accurately might be that the self driving problem becomes much easier to solve.

It will be a brave new world.

970

Cruise has lost $1.4 billion this year

https://www.cnn.com/2022/11/01/business/self-driving-industry-ctrp/index.html

971

This reminds me of the Mobile Telecoms bubble of 2001 when telcos bet the farm building internet portals. It was based on the assumption that millions of new users who would require a full set of internet services on their new smartphones.

Except in 2001 the state of mobile technology was barely 2G, bandwidth limited to a few kbps and supporting a rather pathetic text based cut down version of HTML. It would be years before 3G and decent smartphones would become affordable. There was much talk of ‘first mover advantage’ and telcos engaged in a land grab dedication huge resources to being the first to rollout all the new mobile services. If you had a technology based on SMS/text, it was relabelled mobile internet and huge premiums were paid for what seem rudimentary services.

This tech bubble grew and burst entirely with the telco sector, the players were anxious not to miss out on the huge growth potential of the mobile Internet.

The problem was that the technology simply had not developed sufficiently to justify the huge investments in platforms and services. There was no iPhone in 2001 and no mobile network in place.

Self driving cars sees to be a similar bubble. All these traditional auto manufacturers and new startups desperately trying to capture the huge potential of self driving vehicles.

These companies really ought to know better. They have no shortage of advice to assess the maturity of the technology. But the decisions to bet the farm are presumably made by executives who are more interested in pleasing the investors in financial markets. These people buy reports from companies that make huge predictions about expanding into new markets based imminent technical advances.

Play the fans right and your little tech company can find itself center of attention in a budding war between companies with deep pockets. You need fancy presentations.

I think we are going through a similar bubble with battery tech.

These technologies do eventually develop and become viable. But how long it will take is an exercise in optimism.

With self driving cars the assumption seems to be that many of the problems can be solved in clever software.

So, just add more software developers? Specialists in AI and Big Data. But them some go-faster hardware and sensors. Surely they will strike gold before the competition?

This is incredibly naive.

But I guess there add plenty who know his to ride the wave of investor optimism and get out before the bubble bursts.

Looks like this self driving boom is just about spent and the players are winding down the hype. Something good may come of it in the next decade. Improved safety features, driver assist, maybe some self parking in special parking lots?

972

How do we get from no technology to mature technology except through monetary investment, research and development, and long periods of “not here yet”?

973

Technology isn’t something that gets dropped down like manna every few years. The whole reason 3G is called 3G is because it came after 2G. Smartphones were only possible because of a long series of incremental improvements in earlier mobile tech.

Personally, I think the whole “bubble” narrative is really a wrongheaded way of looking at it. Yes, there are periods where hype outpaces actual process, cycles of greater and lesser investment, stock market crashes, and even the occasional technology that proves to be a total bust. But all of that is just… noise. It ignores the fact that in the background, steady progress is being made on infrastructure, basic technology, manufacturing improvements, and so on. Eventually these things come together and a new product becomes possible. From the outside it may look like it came out of nowhere, but really it depended on everything prior, even the things that looked like failures.

There is nothing on the planet more worthless than the advice of technologists.

974

This brings to mind a short story I once read where people could ‘time travel’ to the future, with the net result being people jumping forward decades to a time when ‘everything is fixed’ leaving fewer and fewer people behind to do the actual work of fixing things.

975

Agree in general. Especially when viewed from the POV of an engineer or technologist. But at the same time …

From the POV of people whose job it is to invest capital for immediate trading profits every day, identifying the places where investor sentiment has gotten ahead of immediate financial return is a vital skill. Both to avoid buying into that, and to be able to foment that same overblown sentiment when talking about what you already own.

Said another way …
There are two games being played here simultaneously. One is technological and legitimately entrepreneurial. The other is financial get-rich-quick wizardry / sophistry / chicanery. A lot of “narrative” is produced by the players (or wannabe players) of the second game. And they own some loud megaphones from which to trumpet their POVs.

976

And many players are in both of the games…

977

Agreed. The recent crypto/NFT stuff being an obvious recent example. My point I suppose is that this stuff eventually gets averaged out in the wash. It’s not a good way of determining long-term trends.

978

We spent a fair amount of time with members of a third, intermediate group (and invested in a couple), legitimate tech builders who were attacking subcomponents in the AV chain with the intent of also creating a lot of narrative and then selling out quickly. So, earnestly technical/entrepreneurial, but also looking to get rich quickly.

979

Cruise is expanding its footprint in San Francisco. As of a couple of months ago they were running three true robotaxis eight hours a day each, presumably they’ll ramp that up.

980

Telecoms companies manage an infrastructure of switched linked wires, fibres and bits of radio spectrum. They make money by selling services on top of this. They develop their infrastructure in a measured way, with new standards agreed and technologies developed by their R&D departments and equipment supplied by a few trusted manufacturers. 2G, 3G, 4G, 5G. They pretty much know what is coming next.

Then along comes the Internet in the early 90’s and an big opportunity to dominate a new market for a mobile version of internet on smartphones,

So they go bat crazy buying whatever moves that they think will give them an advantage over their rivals. Slap down huge amounts of cash on immature technology and run up huge debts. Then, when the customers do not come and the business model appears not to work, they retire bruised and poorer and go into survival mode, they make layoffs, sweat their assets and live off their millions of regular small accounts. The technology and the business models that work eventually emerge years later.

So it is with auto manufacturing. The established manufacturers are facing an existential threat from EVs. On the one hand there is Tesla and their application fast incremental changes to auto manufacturing process that is allowing them to innovate very quickly. On the other hand there is China, which has hundreds new EV auto companies and world leading battery technology and huge export capacity. The driver is the health threat of atmospheric pollution and climate change. Attempts to persuade governments that diesel could be made cleaner and pollute less where proved to be fake.

So an auto industry in crisis and held back by huge sunk costs in traditional assembly lines. They look at Tesla and see it is using powerful computers, cameras and sensors. So they buy into some of that Silicon Valley magic because self driving holds true possibility of big new markets and all that data? Surely there is gold in that!

This leads to huge investments in any company that may be able to deliver these key technologies. Lidar, radar, cameras, fast processors, sensors and communications. Gold rush!

These situations arise from time to time when huge established businesses get spooked when their traditional markets are disrupted and they are forced to respond. They go into survival mode and make radical changes until they again sit comfortably on a steady market and business models that they understand and can manage.

Self driving is an overly ambitious objective. But it has lots of appeal to investors. But the real value is probably in the value if the data the card collect. Not about the road conditions and sensors for self driving but about the performance of the car and the behaviour of the driver. That data can help refine the design and also enable new business models. If you know a lot about how a driver behaves, you can tailor an insurance product accurately. There is a lot of potential there.

EVs are still largely in the personal auto market and are starting to expand into the commercial vehicle market. Commercial driving is a huge part of the traffic on the roads and the companies that own the fleets want to measure their costs accurately. The computer in a vehicle can do a lot of useful bean counting. Fleets of robot trucks driven by AI powered autopilots……that can come later.

981

This sounds interesting. Do you know the name of the story? Or I could ask (if you don’t mind) in the SF story thread in Cafe Society.

982

This is the story I’m Feeling Lucky.