Looks like a route planner problem.
Can you elaborate? The projected path was straight ahead the entire time. Stepping frame-by-frame, the car slowed from 40mph to 25mph by the time it entered the intersection (while the light was still yellow, barely), then braked further to 15mph within the intersection before I overrode it. On the visualizer, the cross-traffic lane guides in the intersection kept flickering between dashed and solid, so maybe it was confused about whether those may have been the correct stopping points for the red light? Or possibly it couldn't accurately gauge the distance to the stopped cars far ahead? (Unlikely; they don't even show up on the visualizer.) I'm not sure it's knowable whether the car would have stopped completely in the intersection had I not overrode it, or whether it would have continued creeping through. My instinct in the moment was not to find out! But I would be very curious to know why it slowed down in the intersection, and why there isn't a higher-level system monitoring for obviously incorrect behavior like this and overriding it (or at least alerting the driver to override). Stopping in intersections for no apparent reason has been a big problem with both 10.4 and 10.5, in my experience.
Nobody does that. Even most humans now rely on car nav or memory. Good 2-D map is needed for FSD. They will slowly add other signs ... but that will only be to override maps - like it happens with speed limit now.
Strong disagree (re "nobody does that"). A good human driver will always look for and obey signs like No Right Turn on Red, especially at non-standard intersections (e.g. not 90°, on a hill, poor cross-visibility, etc.). "Relying on memory" begs the question. And how does a car nav tell a driver not to turn right on red? Ideally the maps would be as up-to-date as possible, but they will always be out of sync somewhere, and the level of driving reliability needed for L4 is much higher than the realistic reliability of maps. Unless of course, the maps are continuously and automatically updated by FSD systems trained to recognize them. But that again begs the question.
End-to-end training is the holy grail of the industry. Everyone talks about it but nobody does it. I think the training compute needs are beyond what is practical currently. May be in 10 years ...
I agree that FSD may not have the compute capability (inference) for true city-street L4 until HW4 at the earliest, perhaps HW5. Eventually it may be retroactively possible to achieve sufficient reliability with HW3 levels of computation, due to advances in ML efficiency, but that will probably be several years after it first becomes possible with HW4/HW5, and more compute will always be better/more reliable.
A problem with the current feature-engineered approach is that it's quite brittle; too much essential and subtle information is discarded during the categorizing and labeling. It may be the only workable approach at HW3 levels of compute, given the current state of ML knowledge, but it still may not get us where we need to be. Perhaps Dojo can leapfrog this and create a system that's even more training-heavy and inference-light, to allow a more end-to-end approach capable of running on HW3? One can hope!