The car is properly recognizing traffic lights pretty darned well, considering the circumstance. It looks like it has a built in understanding that traffic lights are "always" stationary - hence, assigning them static locations on the 3D map - but it keeps having to update the model because the lights are actually moving.
This seems like a very non-obvious edge case that I wouldn't expect an ML team to even consider as a possibility. Now they need to program into the ML model an understanding that traffic lights are typically stationary. Which seems even more difficult to me, from a technical perspective - you don't want false negatives...
The car isn't braking or making any strange maneuvers from what I can tell. I'm actually impressed that it's handling it this well.