Is 10.10.2 the best build so far for me? Yes, it does just about everything better, except makes sometimes moves to far to the left (over a double yellow sometimes) when making a right hand turn.
Are there main adoption blockers? Yes, see below.
Summary - For me the main issues with 10.10.2 are it hesitates and is abrupt when it is safe upsetting other drivers and at harder intersections proceeds when unsafe.
Harshness and upsetting other driver behavior - When the human driver knows it has right if way and the way is clear, FSD has random hesitations, rapid jerks of the steering wheel and sudden stops/accelerations due to lack of confidence at normal intersections, with any level of traffic where it performed remarkably different at the same intersection with the same build several times prior. This makes it very difficult for the human driver to predict how well FSD will navigate a given intersection.
Unsafe maneuvers - Some harder intersections, with obtuse/oblique angles, steep inclines/declines, remarkably different speed limits result in FSD entering traffic without sensing/seeing/reacting incoming vehicles.
Are there ways to fix these main adoption blockers coming in V11? Totally and it would seem that the biggest is using vehicle data to create ground truth for training. I'm referring to
this patent specifically.
How would I go about fixing these issues? While I'm not on the team anymore, I'd specifically look at a way where FSD could have differing levels of trust when entering a new road segment. Currently, FSD has no trust of a new road segment. Each road segment is a black box as the vehicle sees/comprehends new road features. Each visual frame that is fed into the trained model(s) is brand new and has no significance above and beyond the last. I think this could be greatly improved to evolve into each frame having different levels of trust for certain road features. In other words, some of the road features in a given road segment could be immediately be marked as the highest trust level and thus path planning could be optimized to the smoothest, most ideal path possible.
I would define a road segment to be any part of a drivable path that has a given longitudinal length of about or equal to 100m where there is not an intersection. Any intersection is defined as its own road segment with each part of its path being its own segment. Thus a 4 way intersection where each is single lanes has a total of nine segments. 4 that enter, the middle of the intersection and 4 that leave. Each of these segments would have a level of trust associated with it where the highest is used as ground truth and lowest is a black box or no trust whatsoever. Highest trust is established when each visual frame is within a pre-established tolerance range over a period of 3 to 5 frames which allows the system to retain smoothness without impacting safety, even at higher speeds. The level of trust for each road segment starts out at the highest level and then is re-established with every 3 to 5 frames. This is simplistic as tolerances can be dynamically adjusted with several attempts at the same segments overtime. Tolerances can also be saved locally and shared with the fleet for planned routes. I would anticipate shared tolerances route planning data to be less than a megabyte of information as these tolerances could be used as weighted inputs to the real-time trained inference model.
