Originally published at: https://boingboing.net/2020/09/02/a-small-number-of-self-organiz.html
…
That sounds nice in theory, but look at the short video to see what’s happening: three AVs are coordinating together to block all the lanes from anyone behind them overtaking them. Even if in the long run it’s more efficient, we all know what would happen if something like that happened in real life: someone immediately behind the three AV’s will get pissed that the cars are blocking all the lanes not allowing them to pass, and will try and pass on the shoulder, causing an accident.
Yeah it looks like the AI is using each lane equally rather than accounting for the rule of the road that left (or right, depending on what country you are in) lanes are for passing.
Robot shepherds for the American driver. About time.
And the sim shows vehicles staying right next to each other which is a dangerous position, as if intended to cause a traffic slowdown. Clearly this sim is too simple to tell us much.
Maybe, but self-driving cars will always yield to a more aggressive vehicle. Just swerve around them and blast on past! No self-driving car will ever proclaim that traffic laws are for chumps.
Could be done without AV – with hired agents controlling the flow.
But why are we researching how to have better traffic jams instead of how to
- improve teleworking (hello pandemic)
- improve public transport
Yeah, absolutely. It seems that the only prevarication implementation of AV tech will be the equivalent of HOV lanes. Two lanes for traditionally controlled cars and two for AVs. Of course, I don’t see this being practical at all in places where the traffic pattern is changing constantly unless there are dedicated entry/exit points for the lane and they are barricaded from other drivers entering. I would love to be able to sit back and relax while on the road, but there is no way I’d feel comfortable being mixed in with regular traffic.
Try this in Houston and you would have Bill Joe Jim Bob either shooting at you or trying to ram you from behind with his lifted bro-dozer. The amount of hubris and entitlement goes way beyond logic or reason.
Yeah, that’s the kicker. All these agent-based simulations of traffic ignore what happens when aggressive human drivers discover the car ahead is autonomous. And as the autonomous vehicle has to slow down when someone swerves in front of it, you’re back to stop-start congestion.
This should be the first, second, and third response to all AV stories, imho
Road rage would be a hell of a way for an AI to achieve sentience. But I guess Steven King has already thought of that one.
This is why physicists should not be taken seriously when they move above the level of particles. This sim makes all kinds of assumptions that will not bear scrutiny by any behavioral scientist.
- Every car is the same (probably controlled by the same algorithm)
- The road is 3 lanes without entry or exits
- The goal should be maximum throughput/flow
- Humans won’t change their behavior in response to things like this
- Cheaters don’t exist.
- The traffic on the road is always dense but not gridlocked
- Laws will permit this
- Accidents never happen
Any one of these assumptions will probably completely distort their results. A primary goal of roadway construction is to maximize safety over throughput, which is both about exits, signs, and laws (stay right except to pass). People have very different tolerances for speed and danger. This is a lot like the physics models of airplane boarding that fail in real-life, especially when we realize the main goal of airlines is not to board as quickly as possible but to use boarding as a perk and thus maximize revenue. This means boarding a slow as possible wins.
Self driving cars are suuuper advanced AI at this point and may never really turn out (cars may become obsolete before self driving capabilities catch up).
I bet we could do this today:
- All vehicles come with medium range communication systems (say, 10 km range)
- If the in-vehicle AI detects situations likely to lead to stop-and-go traffic jams, the speed calculations above are done
- Instead of full self-driving yadda yadda, a chime sounds and an LED around the perimeter of the speedometer lights up
- Everyone knows if they keep the speedometer needle pointed at the LED, they get to avoid rammy-brakey driving, and can just chill.
- Optionally, cruise control can be given the smarts to adjust its set speed within some limited range, and if the adjustment needed is above that amount then it hands off to the driver as above.
There will always be some aggro drivers who just can’t figure this out, but as demonstrated it doesn’t take all that many informed drivers to influence the others who are just following the traffic speed within the couple of km they can see.
Given the complexity of modern human-driven car computers, it may be possible to just program all new cars to incorporate the algorithms in the article into the onboard computer systems. The trick is making it seem that the car is behaving properly when it’s executing the traffic-clustering algorithm. That’s a user interface issue (admittedly the most difficult part of any computer program).
Some kind of expanded AI assist should be possible, but I think it will be hard to achieve. One of the biggest challenges with both full automatic and assisted driving is handling the transitions to either full manual control or elevated manual control.
People don’t handle these transistions well. These systems are most effective in an immediate, critical situation like collision avoidance – automatically slamming on the brakes to avoid a jack knifed track full of plutonium and ocelots (or maybe something a touch less Hollywood).
But the more you increase opportunities for driver judgment and discretion, the more likely you end up with situations where drivers decide to let the auto run ten seconds too long while they finish reading their text. People start ignoring warnings, they start misjudging threat levels, and that kind of chaotic behavior messes up traffic and can be deadly.
I think there are opportunities to expand on the very basic collision avoidance in operation now, but I suspect it will need to be rolled out very carefully and tested like crazy.
For that matter, we don’t know what people may do when they notice an automated vehicle is behind them.
You also can’t rule out drivers trying to play games with AI vehicle passengers behind them just for kicks, or because they’re mad about something the robot driver did ten miles back.
You could just send information to people’s navigation and ask them to drive in that pattern right now.