Comment by simojo
Comment by simojo 3 days ago
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.
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.