Tell HN: Biggest self-driving challenge is not technology | Hacker News Reader

Tell HN: Biggest self-driving challenge is not technology

2 pointshammock2 comments

In aviation we already have planes that can and regularly do land themselves without any pilot touching the controls. Landing being the most dangerous part of flying (don’t try to compare it to auto-parallel park) It’s called autoland. The plane’s sensors gather left-right and up-down info and then control the plane down into the runway. We do this with planes carrying as many as 350+ souls on board. You would never even know if a pilot was using autoland, and you’ve probably never heard of an autoland accident the way you have with self-driving.

The difference is, for a plane to do an auto-land (Cat III Landing) you need six things. The airport has to be certified, the crew needs to be certified, the aircraft needs to be certified, the company you are working with must be certified with its home civil aviation authority, the airport/country you are doing the cat III must certify the company as well, AND the weather should actually be low vis (meaning other types of landings like visual, VNAV and standard ILS are more difficult/impossible).

If you have just one of these items not certified, then the pilot can’t do an auto-land.

Until we have the same rigor in self-driving, I would not anticipate much success, for safety reasons.

Planes can be automated because their routes are segregated and scheduled (for commercial flights). There is no chance of a pedestrian walking through an air corridor and very little chance of a bicycle being left on a runway. While not perfect, their activities are surveilled by third parties which greatly contributes to system safety.

Monorails, which operate on the ground but not at grade level, can also be automated for the same reasons and have an even better safety profile than airplanes. I'm not suggesting monorails vs cars, but rather pointing out that the environment in which the transit operates is paramount.

Self-driving is trying to automate around the chaos created by human-driven cars on unpredictable routes, while planes and trains have neither of these issues. And that's why it's currently an unsolved problem.

> Until we have the same rigor in self-driving

I think automated robotic self-driving will always need a human-in-the-loop as ironic as that sounds, but for 99.9% of an automated journey, the extra 0.1% needs to be offloaded to humans, for rare edge-cases where the self-driving vehicle has no training data for. For example, random events like road maintenance, wild animals crossing the road, natural events like snow storms, etc.

The people working on this 0.1% metric want to further reduce this percentage and get all the automation to 100%, but you will need humans at some stage. Nature is too random and chaotic for it all to be modeled in a computer system.

Reply on news.ycombinator.com