Flaw in FSD approach?

[insaneoctane]

Does anyone here think that the fact that FSD can't learn your local roadway nuances might be a deficiency? Everything that FSD sees is like the first time and it applies what it "sees" to the NN. Granted, if your nuance happened to make it into the training dataset, it might do well. A perfect example of this is when my car was exiting a parking lot, it turned right onto an uphill slope. It ultimately had to turn right, drive up the incline and quickly get into the left lane to turn left. I've made this maneuver countless times, so I know to turn right into the leftmost lane in preparation of the upcoming light. FSD beta struggled here, it turned right, drove slowly up the incline and subsequently realized that it needed to be in the left lane to turn left at the light, so it ended up diagonally in both lanes, not having enough room to complete the maneuver. Do we really think Dojo and training are going to prepare our vehicles for all of the nuances we encounter everyday?

 

[Tam]

...Dojo and training are going to prepare our vehicles for all of the nuances we encounter everyday?...

Nice theory but what counts is the reality. As CA DMV disengagement reports show, every single Autonomous Vehicle company still suffers disengagement.

 

[JulienW]

Individual car "learning" as it drives sounds VERY sensible and practical (almost anthropomorphic). Having Tesla see "for the first time" the road you drive on to work on is counterintuitive to we humans. Also it would almost certainly be the ONLY way FSD will ever have any chance of reaching my parking spot.

However a counter to this is you CAN'T have some cars learning more complex and different driving patterns than other cars. This would make some Teslas "smarter" and some "dumber" drivers than others. A car in the middle of a city would become a good FSD driver and a car in an easy flat rural area would look like a "country bumpkin" trying to drive in the city. When you advertise your Tesla for sale would you claim it to be in the top 5% of FSDs?

 

[nvx1977]

This part of AI day talks about stitching together the data from multiple cars to get a very accurate reconstruction. I'm assuming this is still in aspirational stages, at least where I live, where there might only be one or two people with FSD beta contributing to the mapping. But it seems like weird local quirks, occlusions, etc should get solved once there's more data for this kind of "hd map."