Hacker News Reader

Comment by spaceribs | Hacker News Reader

Well, I do feel like I've been had after that and a few other videos, so thank you for that.

My point though, is that I don't want self driving cars to defeat simulations or tests, I want self driving cars that can react to real driving conditions.

View on HN

But how else do you imagine scenarios like these are tested? Normally there aren't a lot of people willing to jump randomly out in front of cars to test their reaction.

I really don't know. Even with all the hours of recorded road time by Google/Uber used as a testbed for proper driving conditions, most machine learning systems still suffer from figuring out ways of cheating[1], and in those cases it could be fatal.

[1] "The program even learns how to take advantage of bugs and glitches, like timing jumps so that Mario begins falling again at the exact time that he makes contact with a Goomba. Mario's invincible when he's falling, so the touch kills the Goomba, not Mario, and it gives him a further jump boost." http://www.wired.co.uk/article/super-mario-solved

That's a problem for the test design. Maybe we can develop an approach that uses more randomness to test these systems.

But I dont see how you can avoid having a simulation. These events are just too rare to wait until they occur in real life.

It is not ML deciding where to turn, when to slow down.

I imagine just rolling some objects across the road or some self-riding bicycles should be enough.

But what if the self driving cars fail the simulation? The simulation would be the first thing the car needs to pass, if car passes the simulated, then special tests, then you get approved to test on public roads with a security driver, I have no idea when is the time to test without a driver.

Reply on news.ycombinator.com