-=buzz=-
Member
My take on an explanation/interpretation is this:
AFAIK Goggle/Waymo approach is extensive surround mapping, and a designed algorithm to navigate through this perceived world.
Tesla on the other hand, is using a Neural Network which is learning to drive by itself, based on the input of its sensors.
Therefore, a disengagement for Waymo indicates, that they encountered a situation their algorithm couldn't handle AT ALL. Google will have to extend their algorithm to handle this new situation correctly in the future.
In Teslas case, a disengagement means that the NN just learned not to take this line of action for the given situation EVER AGAIN (Or at least add a significant negative bias to this path). So each disengangement actually improves the AP.
As others have mentioned over time the disengagements should asymptotically approach 0.
If you now imagine all current AP2 cars running in shadow mode, and every beviation between AP and driver behaviour is used as an input for the Neural Net, this should give a massive amount of input from improvment over short periods.
Based on this input from all Tesla drivers, the AP should actually develop an averaged driving behaviour, which means that AP should drive better than 50% of the human drivers. And thats before adding always following traffic laws (no speeding), not getting tired, faster reaction times and having 360° vision.
It should even develop some kind of intuition for dangerous situations, based on drivers easing off the gas or tapping the brakes in certain situations..
AFAIK Goggle/Waymo approach is extensive surround mapping, and a designed algorithm to navigate through this perceived world.
Tesla on the other hand, is using a Neural Network which is learning to drive by itself, based on the input of its sensors.
Therefore, a disengagement for Waymo indicates, that they encountered a situation their algorithm couldn't handle AT ALL. Google will have to extend their algorithm to handle this new situation correctly in the future.
In Teslas case, a disengagement means that the NN just learned not to take this line of action for the given situation EVER AGAIN (Or at least add a significant negative bias to this path). So each disengangement actually improves the AP.
As others have mentioned over time the disengagements should asymptotically approach 0.
If you now imagine all current AP2 cars running in shadow mode, and every beviation between AP and driver behaviour is used as an input for the Neural Net, this should give a massive amount of input from improvment over short periods.
Based on this input from all Tesla drivers, the AP should actually develop an averaged driving behaviour, which means that AP should drive better than 50% of the human drivers. And thats before adding always following traffic laws (no speeding), not getting tired, faster reaction times and having 360° vision.
It should even develop some kind of intuition for dangerous situations, based on drivers easing off the gas or tapping the brakes in certain situations..
Last edited: