Hacker News Reader

Comment by haltingproblem | Hacker News Reader

I find this to be relevant to the self driving problem - Apple/Foxconn could not detect when things had gone wrong on the automation line and stop it, let alone have the line's robots fix it. However, we expect a self driving car to detect when it encounters a novel situation on the road? And it surely will.

If they could not detect it in the confines of a highly controlled factory assembly line (not manufacturing but assembly) then how can a car detect novelty on a cityscape or even a highway?

View on HN

Those are two pretty different domains. In manufacturing automation, some of the things that are hard include gripping objects, handling fluids and pastes, and adapting to design changes.

For self driving cars, the controls are pretty similar, even across car models. Gas, brakes, and steering. It doesn't happen that a supplier ran out of engines and a decision is made to replace the car's propulsion by jet engines, thrust vectoring, or a hovercraft. It doesn't matter too much for self driving cars to be within 0.1 inches of the center of the road, but if your electrical components are offset by 0.1 inches or there's 0.1 fluid ounces too much glue because this batch of glue is more liquid than the previous one, the electronics probably won't work in the end.

And yet if you look at what actual experts, that aren't trying to sell anything are saying, like automotive testing bodies (EuroNCAP, ADAC, FIA, etc) the consensus is that self-driving cars are at least 15 years away. The gap between reality and what salesmen have sold to common people is gigantic.

Also the R&D budget for self-driving cars is bottomless.

There's a limit to the R&D budget for assembling THIS year's iphone, above which it's just better to use humans.

Apple manufacturing budget >> Self driving car r&d budget.

My comment was not about robotics but about detecting when things had gone wrong. Again, the manufacturing lines are in a very controlled environment - loaded with sensors, compute power, perfect lighting, humidity etc. Compare that to the noisy world of roads, traffic, weather and humans.

And in that same vein, why haven't we focused on the infinitely easier realm of rail automation and safety?

We jumped right to the hardest problem set. Probably because it's the most sensational and easiest to get broad financial support by selling people the promise of less rush hour drain.

Less sexy, less money, more government.

The rail automation is solved, entire metro lines are automatic in several places. Unfortunately while it is an easier problem to solve, it's also a less financialy interesting: in a train the labor cost of the driver is much lower than in a taxi.

to be fair, we do have some autonomous rail infra - Metro line 14 in Paris, the DLR in London, the Victoria Line was designed to be autonomous but scuppered by the unions.

The Copenhagen metro is fully driverless (no staff in any role onboard) and consequently runs 24/7.

Replacing the cost of 1 conductor every 200 passengers is not nearly having the same impact. Moreover as others point out, it has actually been done and there are automated lines. Safety advantages are probably slim - when there's a train accident it makes news, because they're RARE.

It's also not going to significantly improve one's choice of transportation - you either have access to rail already, or you don't - regardless of who/what drives the train. Schedules are already tuned and if it was more profitable to put more trains, it would likely happen even with human conductors.

Safety regulations are what prevent more trains from running in many European places. Limiting rail to automated-only trains would allow them to run right up next to each other without the huge gaps between them that exists right now. You could move thousands more people than automated cars could manage for a fraction of the cost.

Is it really?

I thought you generally can scale trains by adding more wagons.

Only for busy point-to-point connections. You can bundle up wagons from different senders or for different recipients into one train for the bulk of the journey, but the first and last leg are often done by truck because it's more cost-effective than running trains with just one or two wagons from/to individual factories.

Now if you had a train network with autonomously driving wagons that self-assemble into convoys, that would be quite something (and probably more cost-effective than truck convoys in the long run because the lower friction of metal-on-metal compared to rubber-on-asphalt).

Yes, there's a safety limit to how close trains can be.

Sure you can make trains bigger but many times trains are already the necessary size for their time windows, destinations, etc.

Also simply making bigger trains puts more control into the hands of a few companies. If there's going to be more competition driving prices down and offering more options, then there needs to be smaller margins in safety controls to make room for more competition.

That works for goods, not so much for people.

2 examples:

1) You can't make the train longer than the station (think of a subway station)

2) I don't need 300 more seats on the 6:30 AM train, I'd like to sleep 30 minutes more and have an option at 7:00 too. The existing 7:30 train arrives too late for my needs.

I wonder if there are companies focusing on solving the opposite problem. In other words, AI to focus on things humans don't perceive (leaving the driving to the human). For instance, figuring out the person in another car is drunk, then alerting you to avoid them. Or detecting emergency personnel needing you to get out of the way (can't tell you how often I see folks blocking or oblivious).

This goes along with my experience in new cars in that the best improvements are those that enhance my ability to drive (such as a backup camera).

The more I work in and watch the ML/AI space, the more I’m convinced the better approach is similar to what you describe: “augmentation” of humans’ abilities and skills rather than replacement of them.

Using advanced AI, or even a bunch of semi-decent models to condense information, highlight things humans might miss, enrich with predictions, etc so that humans don’t have to spend as much time wading through data themselves to try and extract meaning and can instead jump straight to more informed decision making seems like a better approach to me than “lol can we make a neural net that does lawyer things?”

the smaller electronic device is, the higher the precision you need. For such a compact device such as iPhone, the precision is not enough to automate the production yet. For self-driving/flying, there is always some lee-way. The demand of lowering error variance is low.

> not enough to automate the production * yet

* = economically, because low-skilled labor is toooooo cheap

Reply on news.ycombinator.com