Imitation Learning
7 years ago I started comma.ai with a simple idea.
Gather tons of human driving data, state action pairs: (S_t, A_t)
Train a supervised model f(S_t) -> A_t
Drive cars with that model.
The exact original formulation was a model that predicts steering angle from image, then used a PID loop to bring the wheel to that desired angle.
f_steerangle(img_t) -> steerangle_t
This turns out not to work, it couldn’t even drive straight on highways. It would drive for maybe 10 seconds, but then error would ac...
Read more at geohot.github.io