Hacker News Reader

Comment by Methusalah | Hacker News Reader

Sensationalist bs. It'd take some pretty impressive programming to deal with issues like road construction, missing or faded road markings, unmarked dirt roads, and potentially outdated or incorrect maps.

View on HN

I don't think we'll go straight from total human control to total automation directly. I imagine that the first versions of automated cars will require some supervision and only work in certain conditions. This will likely require some sort of system to detect the driver is paying attention, like steering wheel sensors, to prevent people from just going to sleep.

The first versions of automated cars already exist. Mercedes, BMW, and Audi have versions of their cars that maintain a safe distance from the cars in front of you. Mercedes also has a system which will automatically steer the car to keep it in a lane for you. Prius and Lexus have cars that'll automatically parallel park themselves.

I don't think those are programmed into the driving software as much as it has learned them through machine learning. Most of these systems have gone through the equivalent of millions of years of humans driving so I'd trust it over a human driver any day.

Well these systems weren't ever human, so I'm not sure a million human years would help them as much as you hope.

I don't think any of the systems really use machine learning. It's more classical stats and control theory.

Maybe the object recognition stuff uses some trained classifiers, but the higher level systems doing the driving to my knowledge do not.

They're gonna cover 98% of cases. The rest can be picked up be remote operators.

Remote operators work for flying drones, since those don't have to evade children running onto the road with sub-second reaction times.

For cars there is no way electron-based signal transmission technology can control them with an even marginally acceptable latency.

What do you mean by electron-based signal transmission technology? If you're talking remote control, those signals will be photon based. If you're talking internal control, electrons on wires move pretty fast and many cars already utilize them. Is there much latency difference between mechanically closing the throttle with your foot and signaling a servo to close the throttle? Presumably some situations can be analyzed by a processor and trigger the servo before the human controller has even been able to consciously perceive a scene.

One question:

How do you plan to remote control a moving car with photons?

Edit: Huh, ok. I didn't know radio signals are also photons. Still, the speed of light is significant with sub-second decisions. So even for photon transmission remote control of a moving car in live traffic is not feasible.

Latency doesn't matter. What matter is the final outcome. Better skill, software, data etc can compensate that.

Drone pilots can have several seconds of latency. This means remote operators are useless as a backup to the machine if a split second reaction is required.

So a remote operator is going to be given command of a vehicle which has encountered an unusual situation that the automatics have bailed out of?

Ask the pilots of Air France 447 how well that works out (and they weren't even remote). Oh you can't... they're dead.

You still need an algorithm to decide when it's time to handle control to a remote driver. Who's going to handle the edge cases of that algorithm?

Reply on news.ycombinator.com