I encounter this same need in my day job as a controls engineer.
Robots (actual industrial 6-axis robots, not T-800s), PLCs, and computers are surprisingly dumb, but they're dumb reliably, tirelessly, and very fast. Human operators and technicians, in comparisons, are paragons of unimaginable wisdom, but they are slow to react and occasionally make mistakes.
The magic happens when you bring them together correctly.
A human cannot respond in milliseconds under optimal conditions, under poor conditions (such as being fatigued and stressed by hours of unrelenting demand on their attention) they might take seconds to respond. It's possible but you don't really want to relegate basic tasks like lane centering and speed regulation to their subconscious, because they're likely to make mistakes. Computers can do a 1ms PID loop all day.
On the other hand, even with all the massive advancements being made by LLMs these days, a 4-year-old can better understand the gestures and commands from a traffic cop than software that would take a team of engineers and high-speed LIDAR sensors thousands of hours to interpret with a useless 95% accuracy rate.
Let humans do human things and machines do machine things. We don't ask humans to manage fuel mixtures or double-clutching a manual transmission anymore, because the machines can do that better. We shouldn't ask the machines to interpret emergency signage in the future, because humans can do that better.
The optimal level of automation is probably past analog cruise control that's been around for decades and simple suspension caster angles that bias the wheels to keep straight, probably even beyond level 2 radar distance-keeping cruise with automatic emergency braking and automated lane centering, but may never be "punch in an address" automated.
A malicious interaction doesn't need to be a remote drive scenario.
One could see non-removable way points being set in software. Door and window controls locked out. Radio tuned to the most obnoxious playlist possible.
There are numerous ways a malicious interaction could play out, none of which requires 'remote drive' capability.
These are immediate, in the moment, examples with no forethought or planning. Imagine how interesting an actual attacker could make a malicious interaction, or millions of those interactions.
It seems to me that maybe the most efficient course forward, is to spend $0 on computation and just beef up the telecom and pay for a real full-time human driver (not just for tricky situations) in a low labor cost market.
As the story explains, these people aren't driving the car. They're just providing it with high level assistance. "Try moving to here, and then... OK, now we can see there actually is a road, drive to there... and now you're back to a situation you understand".
Latency prohibits "just" remote driving cars, the car needs local smarts to drive itself.
It seems to me that maybe the most efficient course is to have the person who is sitting in the car drive it, thus removing roughly ten billion dependencies, security vulnerabilities, and failure modes.
Sometimes I get impatient at an old person behind the wheel, and my wife says, "He can't see that fast."
There's not a hard line. There's a degradation in response time, and so old people drive slower (at least, they drive slower if they're wise). And then there's a point where even that isn't safe, and they need to stop.
But yes, not everyone can drive. Some are too old. Some are too young. Some can't afford a car for the amount of time they need to use one. Some are dealing with a temporary thing, like a broken leg or a bad sprain, and they can't drive just for a few days, but they still shouldn't drive right now. Some are coming back from a bar, or from a medical procedure involving anesthesia. (I forget the procedure, but I was told "don't sign any legal documents in the next 24 hours" along with "don't drive".) Lots of people can't drive for lots of reasons.
But, you know, we used to have things called "cabs" that helped with all of that. "Not everyone can drive" is orthogonal to "should a car be self-driving, or should a human drive it".
Cabs cost significant money and will continue to cost relatively more and more due to Baumol's cost disease preventing productivity increases (1 driver can only drive 1 cab at a time).
Self driving cars offer the chance for inexpensive point to point transportation. Just because humans were once needed to drive cabs doesn’t mean they always will, any more than we need elevator operators anymore.
We still have cabs, urber... They are all too expensive to be your normal way of getting around. Useful for rare trips or emergencies, but not for the normal day to day trips. A personal car is quickly cheaper, while a cab becomes unaffordable. (Cars are not cheap, but most of the costs are fixed)
The idea of cars being driven remotely by someone on the other side of the world was mentioned in Charles Stross's 2007 science fiction novel Halting State. I think about it sometimes, but I always come to the conclusion that the latency would be too high. A 100ms latency might make the difference between a near-miss and a collision.
But it does offer a reasonable hybrid solution: the software in the car doesn't have to solve all possible scenarios, it just has to be good enough to navigate the common, easy scenarios and should always be able to safely handle any situation for a short window of time, even if "handle" means "come to a quick safe stop within X seconds and turn on the hazards", all the while it's already starting the process of asking for remote assistance to take over whatever tricky situation, hopefully before the local code actually reaches the emergency-stopped case.
We see current cars marketed based on their safety features or crash test results - in this driverless future we'll have luxury brands touting that their remote drivers are based in the USA and can respond 5 times faster than more affordable brands with drivers on the other side of the world.
If you own the car, you'll be the one responding to it. But if self-driving cars work as well as they're hoped to, you won't own the car. Car ownership is absurdly expensive.
I mean, it's absurdly expensive except for all of the other types of car usage right? Companies have been pushing car subscriptions for years because they sound nice: "Don't worry about maintenance, just pick a new car every so often!". Except that it turns out to be even more expensive than just owning a car.
Leasing (unless you have some tax advantage) is typically more expensive than owning.
Ride hailing apps are great, but I don't know anyone who has run the numbers and decided it makes more sense for a daily commute than owning a car (maybe someone with no or extremely expensive parking options?).
I'm skeptical that using a self-driving car service from a profit-seeking company would end up being less expensive than owning your own self-driving car, but maybe we'll see.
Leasing and car rentals have the same problem, which is that you're a single driver hogging an expensive capital asset that you'll use maybe 5-10% of the time. This low capacity factor makes car ownership very expensive. The presumed advantage of self-driving cars is that self-driving vehicles can operate at a higher capacity, factor without incurring the extra costs of needing a human driver to operate them (like taxis) which will make them more useful.
NB: I did not used to think about this much until I recently went to price an extended warranty for my low-mileage car: all of the plans were denominated in years _or_ miles, which meant manufacturers price maintenance in terms of years owned, regardless of usage. The same is true with insurance, which is a huge cost for most drivers: some insurance plans offer a modest discount for extremely low mileage drivers, but to a first approximation the fixed costs of simply owning a parked car are huge. Basically the only saving for owning a parked car is fuel (which can often by offset by the cost of parking, whether that's born by the owner or subsidized by the city.)
Humans are not perfect drivers. They often do make mistakes, but when it is their mistake they think it was a trivial thing, when it was someone else's they think the other guy is bad. Thus everyone thinks they are a great driver - meanwhile millions of people die in car crashes every year (https://www.who.int/news-room/fact-sheets/detail/road-traffi...). computers have the potential do improve a lot on humans, the only question is will they - I'm optimistic but it remains to be seen.
Tesla claims their autopilot is much better than humans - but unbiased sources that should have data are not talking.
Robots (actual industrial 6-axis robots, not T-800s), PLCs, and computers are surprisingly dumb, but they're dumb reliably, tirelessly, and very fast. Human operators and technicians, in comparisons, are paragons of unimaginable wisdom, but they are slow to react and occasionally make mistakes.
The magic happens when you bring them together correctly.
A human cannot respond in milliseconds under optimal conditions, under poor conditions (such as being fatigued and stressed by hours of unrelenting demand on their attention) they might take seconds to respond. It's possible but you don't really want to relegate basic tasks like lane centering and speed regulation to their subconscious, because they're likely to make mistakes. Computers can do a 1ms PID loop all day.
On the other hand, even with all the massive advancements being made by LLMs these days, a 4-year-old can better understand the gestures and commands from a traffic cop than software that would take a team of engineers and high-speed LIDAR sensors thousands of hours to interpret with a useless 95% accuracy rate.
Let humans do human things and machines do machine things. We don't ask humans to manage fuel mixtures or double-clutching a manual transmission anymore, because the machines can do that better. We shouldn't ask the machines to interpret emergency signage in the future, because humans can do that better.
The optimal level of automation is probably past analog cruise control that's been around for decades and simple suspension caster angles that bias the wheels to keep straight, probably even beyond level 2 radar distance-keeping cruise with automatic emergency braking and automated lane centering, but may never be "punch in an address" automated.
reply