One way of putting it: My Apricot Portable came with voice recognition over 30 years ago. But even if I trained it to have 80% success in recognizing my words, that’s not good enough to replace typing.
Five years ago voice recognition systems were at 98% success. But even then, that’s a lot of interruptions where you have to use the keyboard. It’s still not good enough to replace typing.
Siri and Cortana, using cloud computing, have probably double the success rate. 99%. Still not good enough.
And that’s under ideal conditions. A user with an excellent voice, no background noise like other office workers talking, and no technical terms not in the dictionary it’s checking to improve accuracy. For most of us it’s much less than 99%.
Self-driving technology is like that. The testing so far has been under ideal conditions, mostly on the same small, well-3D-mapped, high connectivity roads in good weather. Get outside that, where the car has to do more guessing, and 99% accuracy isn’t good enough. 99.9% isn’t good enough.
They won’t be talking to the millions of cars on the road now. They won’t be talking to the edges of the lanes, hidden under snow. And on a long road trip - even sometimes in the city let alone in the country - they’ll be outside cellular coverage and not talking even to each other.
Again, what the cars can do in ideal conditions is all very nice, but they have to work 100% in non-ideal conditions.
I get this attitude while writing accounting software: I’ll say, “It’s complicated, because we have to handle X.” The customer will say, “Yeah, but X hardly ever happens. Only once a month or so.” And I have to explain, “The fact that X happens AT ALL, means that we have to be able to handle it. It makes no difference whether it’s once a month, once a year or once a minute.”
Back when self-driving cars were a new tech I had an idea I thought would make a great cold-opening for an episode of Law & Order.
1. EXT. HOSPITAL, NIGHT
A lone SECURITY GUARD smokes a cigarette near a curb. A sedan pulls up nearby and comes to a stop, catching his attention.
GUARD: Hey buddy, you can't park there. This is a no-stopping zone.
The guard sees no response from the car and becomes visibly annoyed.
GUARD: Hey buddy, you deaf?
The guard throws down his cigarette, walks brusquely to the side of the car and starts rapping on the window.
GUARD: Hey, I said—
Camera shifts to show CORPSE in driver’s seat with a bullet hole in its forehead. The color of the corpse suggests it has been dead for several hours.
SCENE CUT
The area around the car has been marked with crime scene tape as police officers keep the public back and collect evidence and photographs. Detectives BRISCOE and GREEN are on site.
GREEN: CSI says the John Doe has been dead for at least 14 hours.
The car's an experimental model, for all we know it's been driving this
guy around across three states.
BRISCOE: [Dry witticism about wacky new technologies]
Virtually all cell phone camera lenses are centered on the phone when held vertically. And so the phones are held vertically while filming.
If you want to change that, petition Apple (and Google, to change their guidelines to manufacturers) to put the camera lens half-way down the phone on one side. So that it’s top and centered when the phone is held horizontally.
What? No. My iPhone has the camera in the corner. And anyway, the distance of the lens from the edge of the phone is trivial in relation to the other relevant dimensions when filming other than at macroscopic distances. Vertical video exists because people are stupid, not because they are somehow guided to it by phone configuration. The simple fix would be for the phone to have a square sensor, but for it to default to record landscape footage regardless of how the film-maker chose to hold it. A slightly more sophisticated fix would be for the phone shoot and record square and allow the viewer to choose the crop.
They could coordinate their RADAR and LIDAR data, taking samples in microsecond slices, sharing and assembling it in larger picture than any individual car’s sensors. The larger issue will be how each vehicle trusts sensor data from other vehicles and the security implications around that.
Now you’re vastly complicating things. First the area MUST have great connectivity for the cars to all coordinate with each other.
If the cars are on multiple cellular networks (often interconnected in another city half a continent away) there still must be so little lag that the cars can not only share data will millisecond precision, but coordinate who take RADAR and LIDAR measurements in any given millisecond. With cars constantly entering and leaving the area, joining and leaving the coordination.
Technically possible, but not realistic.
I could see a central control box at a high volume, high speed intersection that each car communicates with directly via a Wi-Fi like system. But even taking all the non-self-driving cars out of the picture - which is at least a couple decades off - I wouldn’t trust my life to it given how many problems laptops have connecting to Wi-Fi. 99 out of 100 cars establishing a connection when entering the intersection isn’t good enough.
Eventually, but it will be a while before self-driving cars become the norm. Yes, most new cars will have it in five years, but it will be another five to ten before most people buy a new car. I doubt it will be cost effective for insurance companies to subsidize everyone getting a new self-driving car, or even retrofitting their existing vehicles.
What irritates me is that self-driving cars have been a trope in SF for some eighty years, but the public is acting like Tesla and Google are really thinking outside the box. I mean kudos to them for engineering it, but someone was going to put this together as soon the components reached the necessary level of sophistication.
Right; even seat belts didn’t become universal overnight. There are still plenty of cars on the road built before they were legally mandated. I have an old motor scooter that doesn’t even have turn signals.
These technologies will be rolled out over time, probably beginning with anti-collision systems that can override steering or brakes in an emergency and gradually getting more advanced from there.
But will we be able to retrofit old classic cars so that we can safely cruise in style? I’ve put a center brake light, rear firewall, seat belts and disc brakes in my 1960 Ford, but a self-driving retrofit would be sweet. That can be done with a kinect sensor and smartphone running open-source software, right?
Or we could just build some true public-trans infrastructure and we could all be sleeping on the way to work. I mean just look at all the people headed in the exact same direction… Even if they are all self-driving it is still massively inefficient.
But what about Waze? It does an app version of pretty much that. All cars communicate with each other about traffic jams, radars etc in real time using GPS plus cellphone data network. Even if there’s blind spots in the network the computer onboard can save data and send it to the network when it finds signal again, helping the drivers that will pass that spot after you.
