- Self driving is not possible
- Self driving is not possible anytime soon
- Self driving is possible, but requires LIDAR
- Self driving is possible, and can be done with normal cameras
I've engaged with a lot of people who presume self-driving reduces onto AGI, therefore it will not be achieved anytime soon if ever. I wonder which of their assumptions is wrong, if this ends up being successful.
Many people assume that a self driving car must be 100% successful. To me it is sufficient to beat the top 20%/bottom 80% of drivers.
So the car will crash if a seagull flies in front of the camera and there is a deer just beyond it. The human would probably crash too.
Computerized safety systems can, should, and are raising the bar, in terms of safety for the top 20% of drivers. "Self-Driving Cars" aren't competing in a static environment, similar tools are raising the safety bar that they have to compete against. This means it will get increasingly difficult to beat non-self-driving vehicles.
And what is the story on driving becoming safer?
It's interesting that you set the threshold there (I presume you meant 'beat the bottom 80% of drivers'). Rationally, we should be happy if their driving performance is above average. I fear that the public (and the courts and the insurers, etc) will require airline levels of per-mile safety and still be wary.
If a human driver kills someone / injures someone / damages property, people get satisfaction or resolution when the human is punished - insurance increases, license points, jail, etc.
But when a self-driving car - even if it's better-than-average - does something wrong, there's nobody to punish and no retribution to exact, so people will be left feeling unsatisfied. That's why the bar is going to need to be so much higher.
Pick something you pay extra attention to. Now imagine being informed you can't do that any more, you have to take what [this robot] does for you. Also remember all the terrible UI changes in software that have been done for a majority instead of your use-case.
Reluctance until it's substantially better seems reasonable to me. Somewhat unfortunate on a species level, but not for many individuals.
That is still a pretty high ask for current software.
Those two obstacles won't be overcome until AGI. At best their frequency will be brought down a bit but not by nearly enough orders of magnitude.
Also even Tesla handles the obstacle in the road situation well. Their FSD beta will happily enter the left lane to go around obstacles.
If an AI system's default response is "come to a stop safely" then it's going to be way ahead of a lot of human "unexpected situation" handling in cars.
Yes, people are bad at driving, because they don't pay attention, panic, make mistakes etc. But ML models tend to freak out at slight variations on mundane circumstances; a cyclist crossing the road at just the right angle and the wrong colour of bike, that sort of thing. The thing self driving cars need to avoid is killing people in broad daylight for no discernable reason, and that seems like the kind of thing that you'd need a mind for. It's the same issue as with adversarial image manipulation to fool image recognition; if changing 3 pixels can turn a frog into a toaster, you aren't really "seeing" the frog at all in a symbolic way, and not seeing a road symbolically seems like a recipe for disaster.
This, I think, is the thing that people miss when they say "self-driving cars don't need to be perfect, they just need to be better than human-drivers, who aren't actually all that great".
From a public confidence perspective, it doesn't matter if a self-driving car crashes one tenth, one one-hundredth as often as human drivers; as soon as you see a self-driving car kill someone in a situation that a human driver obviously would have avoided (like in the adversarial image kind of scenario), you've totally destroyed any and all confidence in this car's driving ability, because "I would never, ever have crashed there."
Sure it's probably feasible for cars to handle hand signals in the happy path, but anything outside of that will be disastrous. How will the car understand and communicate with a person who doesn't use the standard signals, aside from having some level of intelligence?
Sometimes. Other times they confusingly gesticulate or just shout out things, or even give conflicting signals. Humans can interpret these without much effort but it's a hard AI problem.
I don't expect that we will ever see the day where everyone is OK with self-driving cars on the roads regardless of the safety statistics, because there will always be edge cases of crashes and personal preferences around driving styles.
We have the advantage of retrospect when looking at these claims which now may seem more possible then before due to our better understanding of the difficulties and technologies to address them.
I think many people confuse their excitement for the promise of having self driving cars and the actual technical and political barriers that still need to be addressed to bring this to reality (ranging from robustness, infinite number of edge cases, perception, insurance, or public policies).
I think most of the confusion is due to deceptive marketing. Sales people like Elon Musk have been saying that the big dream of true self driving is just around the corner for years now, even though they were nowhere close.
Define “soon”. For me it means 50 years. That’s an infinitesimally small amount of time. If we can have level 5 autonomy in 50 years I will say that progress was rapid and we really excelled.
> The first working device to be built was a point-contact transistor invented in 1947 by American physicists John Bardeen and Walter Brattain while working under William Shockley at Bell Labs. The three shared the 1956 Nobel Prize in Physics for their achievement.
Given sufficient qualitative and quantitative investment self driving may be technologically feasible at some point in the future but not price competitive with human driving.
Self driving proponents seem to assume the tech is free. Just keeping sub meter scale 3d street maps up to date has a massive cost.
> Just keeping sub meter scale 3d street maps up to date has a massive cost.
True, but probably not actually necessary.
Well, for one thing, I don't want anyone in populated parts of the US to live on $10 an hour, not even a cab driver. What happens when we double that? Cost of living goes up over time, but cost of technology goes down. When do we reach the tipping point? Have we already?
If wages go up to $20 an hour, then maybe there is general inflation and robotics goes up to $200 an hour. In the world today we have cheap labor and expensive energy. That is why automation is not replacing humans. Look into the history of the British industrial revolution.
Please describe the last factory you were in.
Here is the reality that self driving will confront:
cost effective, reliable, competitive automation is hard, even for seemingly closed mundane taskes like palatising products.
Yes, it's expensive, however, as we refine algorithms, and the processing systems, it gets cheaper, especially amortized over more cars, where the per-car cost becomes affordable.
ML doesn't work well enough for offline maps generation either, and all the high quality maps require human editors for final touch.
All this work is currently done because realtime perception doesn't work well enough, and you can have a much more reliable system with the aid of the maps. Having a 3D base map of the world makes the realtime perception problem far simpler, and it makes fancy sensors less critical.
In the US, where the cost of labor is very expensive, self driving will make sense, even if it's expensive, but someplace like China or India, where a middle class person can afford a driver, it probably makes less sense, though the push for it in China is probably the strongest that I've seen anywhere in the world.
Self driving does not scale at all now. Because as you said these special maps need to a lot of human labor to make and the cars need a lot of sensors on top of the auto patform. Labor is not even the main cost driver in person transportation.
Why are efforts not focused primarily on interstate/highway travel, specifically, collaboration with DoT for mesh/distributed/coordinated long-term travel?
I don't need a self-driving taxi. I would pay 20K for a car which participated in a federally-regulated framework which let me let go of the wheel when I get on a highway and let teh emergent cloud determine how best to move my car and all the others on dedicated/reserved lanes in coordinated "trains" for hours.
The wins here seem like no-brainers. Sidestep all jurisdictional nonsense; optimize commerce and personal travel; automatically handle emergency vehicles and other unusual conditions; etc ad infinitum.
All my car needs to be able to do other than existing lower-tier self-driving/driver assist basics, is join the borg.
Coordination for traffic flow management seems like an unbelievable win.
But no, all we seem to be getting is cyclist-terminating taxis which cost $250K each and are, IMO, doomed in target-rich environments like SF to not forseeably adequately the last 8% of anomalous novel cases.
Just don't get it.
For 95% of people I think the value is in letting them do other stuff while driving long distances/times instead of stuck behind the wheel. This seems many orders of magnitude easier than trying to tackle inner city driving.
Other car companies are a few years behind, but even something as simple as adaptive cruise control + lane centering is a huge help on freeways.
Nobody is saying that self-driving cars are perfectly safe. Considering how many Teslas are on the road and how many miles are driven on autopilot, it would be surprising if there weren't any deaths. As long as it's safer than unaided human drivers (which it is), it's a net win.
https://www.motortrend.com/reviews/cadillac-super-cruise-is-...