Hacker News Reader

Comment by treis | Hacker News Reader

The problem is that as soon as you define a test then these companies will put an emphasis on making sure their cars work for the test instead of the real world. Then, when something happens in the real world they say well it passed the driving test.

It's one of the perpetual problems of regulation. As soon as you define a test or requirement companies will start gaming the system to meet them instead of addressing the purpose behind the regulation.

View on HN

It still makes sense to have a test. We put every citizen through a driving test. Those are gamed too, but at least they learned a lot of basics... and they are probably better drivers than people not required to take a test. After that, market forces will hopefully go the rest of the mile.

I mean heck, this latest incident wouldn't have passed the original Darpa challenge that got us here.

Pretty sure that driving tests for humans do absolutely nothing - if you eliminated the whole process (and replaced it with a simple age requirement) it wouldn't show up in safety stats at all

On the other hand, any self driving company is running an extremely complicated and extensive test suite on their code. Developing these tests is a core engineering challenge to making a sdv and any test developed by a regulator will be a joke by comparison (unless we give the regulator billions of dollars to essentially replicate the work.)

I agree that the USA driving test system is insanely weak, but they definitely do something. I failed my driving test twice before passing it when I was a college kid in Arkansas. One thing I noticed is the examiners seemed to intentionally try to yell and stress me out. In retrospect, that was a smart way to ensure only someone who has become comfortable enough with driving that they have extra processing power left to handle dealing with passengers is able to become an unsupervised driver. It's a low bar, but it's still higher than "that's the brake pedal to stop, that's the gas pedal to go, have fun"

While I understand that could apply to north america, I doubt it would apply to Germany. You only get a license here if you need a car in the near future, as it is no longer a 3-digit cost.

You can have some required tests but those tests will be a part of the approval process, not all that pass the test get the approval but anyone that fails the test goes hope in shame.

The test could have basic things like a car stopped in the middle of the road, a fake person that is fallen on the ground, a tree on the road, a big stone on the road, a big hole in the road and other similar tests.

You can change the obstacle types and sizes so tests differ. The company will not be allowed to use a different software fort this tests so at least you can test that the sensors are good enough and that the software can handle basic things.

This tests would be like the medical tests a driver must pass before starting the driving school, we can trust Uber like companies to perform this tests inside before putting the cars on the road for public road testing.

If the tests are too basic I am sure clever people can find even better tests to stress the sensors and software before allowing the dangerous tests on public roads.

It could be a moving target.

Maybe you could start with a generic test, for instance: x miles in random location with random conditions.

After that you could add tests when a problem appear. For instance, you could add the conditions where that woman was killed the other day. When another accident happens you would add a test appropriated to that conditions.

Regression testing for driving.

There have been auto crash safety tests in the auto industry for several decades now. I imagine it would simply be an extension of that testing. Yes -- the companies will put emphasis on making their cars work for the test, but just like crash testing some companies will most likely distinguish themselves by designing above and beyond the standard test.

The point is to have a documented standard to compare automated systems by.

I think we're all better off due to automobile crash testing, who's to say that optimizing for tests does not also benefit real-world scenarios?

Reply on news.ycombinator.com