Hacker News Reader

Comment by simojo | Hacker News Reader

I'm curious as to what kind of control stack Waymo uses for their vehicles. Obviously their perception stack has to be based off of trained models, but I'm curious if their controllers have any formal guarantees under certain conditions, and if the child walking out was within that formal set of parameters (e.g. velocity, distance to obstacle) or if it violated that, making their control stack switch to some other "panic" controller.

This will continue to be the debate—whether human performance would have exceeded that of the autonomous system.

View on HN

From a purely stats pov, in situations where the confusion matrix is very asymmetric in terms of what we care about (false negatives are extra bad), you generally want multiple uncorrelated mechanisms, and simply require that only one flips before deciding to stop. All would have to fail simultaneously to not brake, which becomes vanishingly unlikely (p^n) with multiple mechanisms assuming uncorrelated errors. This is why I love the concept of Lidar and optical together.

The true self-driving trolley problem. How many rear-end collisions and riders' annoyance caused by phantom braking a manufacturer (or a society) is going to tolerate to save one child per N million miles?

Uncorrelated approach improves sensitivity at the cost of specificity. Early sensor fusion might improve both (maybe at the cost of somewhat lesser sensitivity).

With above-average human reflexes, the kid would have been hit at 14mph instead of 6mph.

About 5x more kinetic energy.

But would a human be driving at 17 in a school zone during drop off hours? Id argue a human may be slower exactly because of this scenario

> would a human be driving at 17 in a school zone during drop off hours?

In my experience in California, always and yes.

Maybe we should not only replace the unsafe humans with robots, but also have the robots drive in a safe manner near schools rather than replicating the unsafe human behavior?

One argument for the robots is that they can be programmed to drive safer, while humans cant.

But that depends on reliability, especially in unforseen (and untrained-upon) circumstances. We'll have to see how they do, but they have been doing better than expected

Depends on the school zone. The tech school near me is in a 50 zone and they don't even turn on the "20 when flashing" signs because if you're gonna walk there, you're gonna come in via residential side streets in the back and the school itself is way back off the road. The other school near me is downtown and you wouldn't be able to go 17 even if you wanted to.

Yeah, if a human made the same mistakes as the Waymo driving too fast near the school, then they would have hurt the kid much worse than the Waymo did.

So if we're going to have cars drive irresponsibly fast near schools, it's better that they be piloted by robots.

But there may be a better solution...

Kinetic energy is a bad metric. Acceleration is what splats people.

Jumping out of a plane wearing a parachute vs jumping off a building without one.

But acceleration is hard to calculate without knowing time or distance (assuming it's even linear) and you don't get that exponent over velocity yielding you a big number that's great for heartstring grabbing and appealing to emotion hence why nobody ever uses it.

Reply on news.ycombinator.com