Man trapped inside driverless car as it spins in circles

[Zero999]

Really, the best way forward with self-driving cars is to mimic human intelligence and take advantage of two features humans do not have but are very simple (low hanging fruits) on machines, namely (1) lack of fatigue and random erroneous actions (against the algorithm) and (2) supernatural* physical skills like having ten eyes instead of two, lidars (sensing distance with light), radars (sensing metal objects with radio frequencies), and so on.

*) from human perspective

One thing which Tesla is clearly doing wrong IMHO is they insist on neural network processing of input data not much different from humans (camera-only). I mean, that can be made work, but it misses the opportunity of doing something better and more easily. But it could also be just a public decoy, it is well possible Tesla is seriously working e.g. on low-cost solid state mass-manufacturable LIDAR. I would if I were Elon Musk, and I might also be quiet about it, and make all the copycats put all their money into training NNs.

I still remain a bit cynical over driverless cars and other road vehicles. No doubt it's something which will be technically possible, but whether or not it will happen is another thing. The technology has existed for a long time to fully automate railways but nearly all locomotives still have human drivers.

Machine learning is generally a no go for safety critical systems, because of the black box nature of the system. It's impossible to predict what it's going to do, with a given set of inputs. I suppose it can be done if it's operating on top of another hard-code supervisory control system which prevents it from doing anything dangerous.

[tszaboo]

To put it shortly, stopping safely isn't significantly easier than driving safely, and any system for it won't be significantly smaller and more auditable.

I expect every vendor to simply put this task on the main self-driving system, ever time. Sorry to disappoint

Stopping is a Level 2 feature of self driving, antonymous driving is level 5. The complexity of software is orders of magnitude different.

[NE666]

I still remain a bit cynical over driverless cars and other road vehicles

I won't put words in your mouth but perhaps that's in no small part down to us both being in the UK?

I see news images of Tesla, Google et al with their LIDAR etc. equipped experimental vehicles gliding along the wide, six-lane, concrete roads of California. And I try to mentally map that onto the, essentially upgraded cart tracks, which make up so much of the UK's road network outside of London and the major cities.  Perhaps I'm loosing some plasticity of thought with age , but I cannot reconcile the hugely multivariate scenario of driving country lanes and B roads, with so very many errant possibilities, with a 'super algorithm' that can really be expected to perform at the level of a competent human driver, or better.

[tom66]

Waymo can drive in downtown SF bay with double-parked cars and pedestrians on suicide missions, and Tesla FSD can drive on some seriously tight roads in the Berkeley Hills, so I wouldn't rule out European-type FSD just yet.

[jonovid]

IMO driverless cars should be confined zones or areas that are entirely mapped for them.
avoiding getting lost in unfamiliar circumstances.

[NE666]

Waymo can drive in downtown SF bay with double-parked cars and pedestrians on suicide missions, and Tesla FSD can drive on some seriously tight roads in the Berkeley Hills, so I wouldn't rule out European-type FSD just yet.

Indeed, I'll never rule it out completely. After all, they said man wouldn't fly. And there are situations and conditions under which it would probably fair much better than the average human driver, such as the congested motorways of London and the home counties.

However, those examples above are still a world apart from a single-width country lane in deepest, darkest East Anglia. Add inclement weather (e.g. floods, fording water) or seasonal traffic (e.g. combine harvesters) and the equation is very different. Notice how in the Tesla FSD example, there's really no other traffic to contend with, let alone something oncoming which completely blocks the road. I just can't see fleets of Jonny Cabs being able to handle that without causing chaos, at least, not this decade.

[Siwastaja]

I have witnessed a completely mentally capable (not drunk) human driver slowly (walking speed) drive directly into a ditch, missing the intersection by 10 meters, simply because of a smooth and even layer of pure, fresh white snow everywhere and overcast but still pretty bright (thin layer of clouds) day giving extremely even lighting. Human brain's vision system simply could not do the required 3D mapping from the stereo images captured by eyes because everything is evenly colored and the shape of the ditch blurred. At the time I was developing "solid state lidar" (or more accurately, 3D time-of-flight) stuff and it occurred me how funny it is that the 10€ sensor we were using would trivially see the ditch. (At the same time, Tesla released "stunning" demos of their camera system producing point clouds, and I immediately spotted how the road surface itself was completely missing from the point clouds, which is understandable since it's a smooth gray surface. But I'm sure they are doing better already.)

It is also surprising how many human drivers have serious troubles staying on their own lanes on narrow roads. I mean, of course, with no upcoming traffic of course it's normal to slightly deviate from the lane if the road is narrow enough. But what surprises me time after time are those who keep driving in the middle of the road even with traffic to opposite direction. When two with the same sort of brain damage cross each other, an accident is very close.

Also very common pattern seen here is completely ignoring pedestrians and cyclists. Rules about this are very clear, they are road users just like cars so you blink and overtake them as you would overtake another car, giving plenty of physical separation (something around 2 meters, exact number here is irrelevant). If you can't due to other traffic, you slow down and wait for a safe occasion. Instead, many people just drive some tens of centimeters from pedestrians and cyclists at 100km/h simply because they are not interested about following the rules.

So while narrow roads and difficult weather are very difficult to self-driving cars, they surely are difficult to human drivers too.

Snow specifically makes it difficult to spot where the road is; human only has non-calibrated pair of eyes not "designed" for photogrammetry applications, with a pretty powerful neural network system which works surprisingly well for the crap input it gets, but sometimes fails miserably.

I would definitely add multi-source sensor fusion to that mix, to have better chances of knowing where the road is (or even better, where my lane is) in presence of a lot of snow. As a human I still have a pretty powerful brain, but I can't connect to satellites for positioning, and I cannot shoot lasers and measure distances with my eyes, and I also cannot send and receive radio waves. Not even ultrasonic like bats.

[NE666]

It is also surprising how many human drivers have serious troubles staying on their own lanes on narrow roads. I mean, of course, with no upcoming traffic of course it's normal to slightly deviate from the lane if the road is narrow enough. But what surprises me time after time are those who keep driving in the middle of the road even with traffic to opposite direction. When two with the same sort of brain damage cross each other, an accident is very close.

Then it might surprise you to learn that this practice was (maybe still is, it's 20+ years since I last attended such instruction) actually taught on UK "Advanced Driving" courses, as part of "defensive" driving skills. The idea being that it gives you extra visibility around the next turn in the road, and extra time to react.  However, clearly it has to be exercised with caution and full attention to the road - being on the wrong side of the road as you turn a sharp, blind bend isn't what it intended here and is just plain stupidity.  It also means that you are further out from road junctions, should someone joining the road you are travelling on not see you in time, and start to pull out onto your side of the road. Ditto animals etc. running into the road from the edges.

[Zero999]

Waymo can drive in downtown SF bay with double-parked cars and pedestrians on suicide missions, and Tesla FSD can drive on some seriously tight roads in the Berkeley Hills, so I wouldn't rule out European-type FSD just yet.

Again, it's not about whether it's possible, but whether it'll happen on a wide scale. Having the technology is one thing. Widely adopting it is another matter. There are many things which can theoretically be automated with decades old technology,  railways as I mentioned above, but in reality most still have human operators.

[NE666]

There are many things which can theoretically be automated with decades old technology,  railways as I mentioned above, but in reality most still have human operators.

In the UK at least, I think that's in no small part down to the powerful rail unions. I can't imagine them 'allowing' even trials of mainline full automation to be performed, not without mass industrial action brining rail travel to its knees.

However, it's probably also true at this time that using trained human drivers is still cheaper than creating, proving, operating and maintaining a fully autonomous network.

[Zero999]

There are many things which can theoretically be automated with decades old technology,  railways as I mentioned above, but in reality most still have human operators.

In the UK at least, I think that's in no small part down to the powerful rail unions. I can't imagine them 'allowing' even trials of mainline full automation to be performed, not without mass industrial action brining rail travel to its knees.

However, it's probably also true at this time that using trained human drivers is still cheaper than creating, proving, operating and maintaining a fully autonomous network.

It's not the same the world over though. There are other countries where unions aren't such a problem.

[Marco]

Waymo has remote controllers to get it out of trouble, trouble which it gets itself into frequently. In urban driving just coming to a full stop till remote controllers get it unstuck is an acceptable and expected failure mode. Obviously doesn't work on the high way, which is why the only level 3 highway systems follow the duckling.

These are fundamental limitations. To get rid of remote control and do better than follow the duckling requires true AI and complete trust in that AI. No more fragile rules check. AI gets to do whatever the hell it wants, because every rule needs it exceptions and those exceptions can't be predicted ahead of time.

Of course when we get true AI there will be bigger changes than how we get around, it all becomes a bit moot.

[tom66]

I think the issue with train automation is the unions for one (they would never tolerate self-driving train tests on the same lines their members use), but also the consequence of failure.  In the UK, we already tolerate around 1,500 deaths per year due to cars, around 5 billion deaths per billion km.  Self-driving cars might bring this down to 0.5 deaths per billion km (hypothetically).  Rail deaths are currently around 0.3 per billion km, so rail is already around 16x safer.  Additionally, rail does not have the high labour cost that taxis do per passenger.

It is not clear that eliminating a driver, who receives a salary of ca. £65,000 per year (in the UK), will significantly make train travel cheaper or more accessible.  That train rain driver costs nominally ~£50 per hour after employer taxes and benefits; a single train ticket frequently costs around that amount, and a train can carry hundreds of passengers.  Most issues with railways are down to capacity constraints rather than labour constraints, there simply isn't enough space for all of the trains we want to run, and there aren't enough conveniently-located stations to make train travel competitive with cars in most circumstances.

[Marco]

Train drivers can also effectively supervise ... unlike drivers of cars.

The level of annoyance necessary to pay attention with Level 2 automation in cars is unmarketable. Which is why the systems are permissive and the cars keep needlessly running into things with level 2. Train drivers are paid to live with extremely annoying systems to test their attention.

[Zero999]

I think the issue with train automation is the unions for one (they would never tolerate self-driving train tests on the same lines their members use), but also the consequence of failure.  In the UK, we already tolerate around 1,500 deaths per year due to cars, around 5 billion deaths per billion km.  Self-driving cars might bring this down to 0.5 deaths per billion km (hypothetically).  Rail deaths are currently around 0.3 per billion km, so rail is already around 16x safer.  Additionally, rail does not have the high labour cost that taxis do per passenger.

It is not clear that eliminating a driver, who receives a salary of ca. £65,000 per year (in the UK), will significantly make train travel cheaper or more accessible.  That train rain driver costs nominally ~£50 per hour after employer taxes and benefits; a single train ticket frequently costs around that amount, and a train can carry hundreds of passengers.  Most issues with railways are down to capacity constraints rather than labour constraints, there simply isn't enough space for all of the trains we want to run, and there aren't enough conveniently-located stations to make train travel competitive with cars in most circumstances.

On the other hand, the unions are a very good argument in favour of automation: no strikes, staff shortages and the ability to run 24/7. You're right that implementing it is the tricky thing. Past governments have been firm with unions, so it's not impossible.

[EEVblog]

I still remain a bit cynical over driverless cars and other road vehicles. No doubt it's something which will be technically possible, but whether or not it will happen is another thing. The technology has existed for a long time to fully automate railways but nearly all locomotives still have human drivers.

Machine learning is generally a no go for safety critical systems, because of the black box nature of the system. It's impossible to predict what it's going to do, with a given set of inputs. I suppose it can be done if it's operating on top of another hard-code supervisory control system which prevents it from doing anything dangerous.

That's been my opinion for the last decade and remains so.
It's not necessarily a technology problem (although plenty remains, perhaps impossibly so in some regard, e.g. AGI), it's a human problem and political problem.
All it takes is one big high profile deadly crash to have full self driving banned in that state/country. Stats will not matter, human emotion rules policy.
Unless they are getting a huge bribe, no politician is going to want to put their arse on the line and approve full self driving on all roads.

[Analog Kid]

Nice segue into my take on all this, which is that there's a huge element of unfairness in this whole self-driving mess.

Which is this: self-driving is being pushed by the technophiles and technocrats, who never cease to be fascinated by the latest and greatest Shiny Object, which nowadays is Self-Driving Cars, powered, natch, by AI.

Find and dandy: this is all being driven by their unshakable faith in technology.

However, there are a lot of other folks out there who don't share that Panglossian view of things, and who moreover are quite skeptical of the claims made by the self-driving pushers.

You could say that this is just a difference of opinion, or maybe even go further and chalk up the resistance to self-driving to ignorance of technology. However, the element of unfairness is this: whoever wins this tug-of-war is going to set the agenda for transporation for the foreseeable future. Which means that if they (the technocrats) win, those of us who have grave doubts about the wisdom of going that route will be overruled. Meaning that we will now be living in a world that we would never have chosen, if we had had that choice.

Which, of course, is the way most of these decisions have always been made here on planet Earth, and will probably continue to be made.

[Njk]

AFAIK, a DARPA-sponsored race of a true self-driving vehicles is an annual event somewhere in the US

[Analog Kid]

AFAIK, a DARPA-sponsored race of a true self-driving vehicles is an annual event somewhere in the US

Oh, goody; spectators get a twofer, a race and a demolition derby!

[tom66]

Self-driving cars represent one part of the future of high automation and low employment.  Driving a truck is the most common job in 31 out of 50 US states.  The idea that ca. 5-10% of all employment could evaporate in a decade or thereabouts if the wild growth of companies like Waymo is anything to go by should be making people worry and we need to be looking at alternatives to 100% employment.

I could be wrong and just like horse farriers were replaced by car mechanics, it could turn out that automation creates yet more jobs, but it is not yet obvious how that will happen, so we at least need to have some idea of what the path looks like in a post full-employment society. That might be UBI, or it might be state-owned housing (both sounding quite socialist to many), or it might be outright anti-capitalist revolution with the self-driving cars being burned in the streets so the taxi drivers still have a job . 

[Zero999]

Self-driving cars represent one part of the future of high automation and low employment.  Driving a truck is the most common job in 31 out of 50 US states.  The idea that ca. 5-10% of all employment could evaporate in a decade or thereabouts if the wild growth of companies like Waymo is anything to go by should be making people worry and we need to be looking at alternatives to 100% employment.

I could be wrong and just like horse farriers were replaced by car mechanics, it could turn out that automation creates yet more jobs, but it is not yet obvious how that will happen, so we at least need to have some idea of what the path looks like in a post full-employment society. That might be UBI, or it might be state-owned housing (both sounding quite socialist to many), or it might be outright anti-capitalist revolution with the self-driving cars being burned in the streets so the taxi drivers still have a job .

Again, just because it's technically possible, it doesn't mean it'll happen.

There's a far greater chance of mechanics losing their jobs, since EVs are mechanically much simpler. Lots of it is just plug a box in to diagnose faults than doing any real work.

[Siwastaja]

There's a far greater chance of mechanics losing their jobs, since EVs are mechanically much simpler. Lots of it is just plug a box in to diagnose faults than doing any real work.

That happened already 40-20 years ago. Per kilometers driven, car repair and maintenance was a big thing in 1960's, 1970's, still in 80's, like what is now known as "gas station" was back then "service station", even the smaller ones employed a car mechanic. This job started disappearing already from late 1980's due to much improved reliability (and increasing complexity; an interesting combination difficult to grasp for many) of cars and finalized in 2000's by pretty reliable but nearly irrepairable cars. Sure, even further loss of this job is possible but not significantly anymore I think. Running specialized emissions analysis computers is replaced by running specialized EV battery analysis computers. Engine oil swaps become less frequent but normal checks to brake system, suspension parts etc. and replacing worn parts remain.

But in the end, certain jobs disappearing is normal since the beginning of times. We don't have cave painting artists and mammoth hunters and horse carriage drivers and horse caretakers (feeder people, what is that called anyway) anymore. Buses don't have a separate worker collecting tickets, it has been automated long time ago too. People need to adjust and invent new jobs for themselves. Expecting someone to magically give you a job and maintain it until you die is pretty much socialism.

[schmitt trigger]

Which is this: self-driving is being pushed by the technophiles and technocrats, who never cease to be fascinated by the latest and greatest Shiny Object, which nowadays is Self-Driving Cars, powered, natch, by AI.

That, and by C-suite types of the transportation companies, who have wet dreams of a business which is not dependent of those pesky human beings. Which tend to get sick, require humane working conditions and have absenteeism issues.

[tom66]

Again, just because it's technically possible, it doesn't mean it'll happen.

There's a far greater chance of mechanics losing their jobs, since EVs are mechanically much simpler. Lots of it is just plug a box in to diagnose faults than doing any real work.

Predicting the future is fraught with difficulty, but we can look at market economics.

Waymo is currently charging around 70 cent per mile for their SDC taxis in Phoenix, AZ.  The IRS allows drivers to claim back around 70 cent per mile so it is fair to assume that the average car costs about that much to maintain once depreciation, insurance and maintenance are factored in.   Probably a bit more.

Now imagine Waymo offers something like a $400 per month subscription for unlimited SDC usage ("fair usage policy applies" blah blah blah).  Now people are going to start looking at their car lease agreement of $300 per month which doesn't include fuel and insurance, and go, hang on, why would I own a car?

The network effect will start to take over, when you have so many taxis available that you can hail one in 30 seconds.  That is not currently feasible for human-driven taxis for a few reasons:
- Drivers need a wage, and there's only so much business for a $3-4 per mile taxi (basically, the people who really need to go somewhere; and the people wealthy enough to not care.)
- Drivers need to be paid regardless of whether there's a customer in the car.  A taxi has to go ~200% the distance it delivers a customer because of the return journey.  Once there are more customers, this amount reduces, but the problem is the number of customers are limited by the scaling which is limited by the price.

In theory, if universal self driving is cracked (big "if"!) the cost per mile will be comparable to maintenance, depreciation and electricity cost, plus a profit margin, but it'll be almost always cheaper to take a taxi than your own car.

I think once we get to that point the transition is inevitable.  Just like in the 1900s you probably couldn't predict the rise of the motor car - it went from a toy for rich people to a necessary tool in 20-30 years.   Or for a more recent example: imagine being on the board of Nokia or Blackberry in 2006 and you hear a rumour that Apple is going to launch a phone.  "Apple?  That Apple?  The computer company?!  Good luck to them, they'll never figure out the software.  We have the edge there, customers won't want new and untested..."

[coppice]

There's a far greater chance of mechanics losing their jobs, since EVs are mechanically much simpler. Lots of it is just plug a box in to diagnose faults than doing any real work.

EVs basically replace an ICE engine, gear box and fuel tank with an electric motor and battery. ICE engines and gearboxes have been migrating towards throw away when it fails items for the last 2 decades. Still mechanics do quite a lot of work replacing worn bushes, brake pads and other things common to EVs and ICE cars. In fact, allthoughthe brakes in an EV see less wear and tear, they suffer greater rusting problems than ICE cars. I'm not sure EVs represent a further loss of mechanic work over that already lost in the ICE world.