Have these researchers seen Tesla's latest FSD version 12.3, which uses neural nets instead of hard-coded instructions? For example this video: https://twitter.com/WholeMarsBlog/status/1768714246521827643 "Tesla Full Self-Driving Beta 12.3 Drives 85 Minutes from Palo Alto to San Francisco during Rush Hour with Zero Interventions". I was also a skeptic until recently.
One of these times, it will be, but I'm going to bet against it and be right most of the time.
At 12:56, the vehicle drives through a crosswalk with pedestrians present, but misses them.[2] A similar situation appears near the very end. This system drives through crosswalks rather aggressively.
There are no hard cases. No heavy traffic. No construction. No emergency vehicles. No double-parked cars. Almost always, either there's plenty of open space ahead or another car to follow. When heavy traffic is encountered, the human driver changes the route. This looks like a cherry-picked try from many trial runs. It's way below Waymo level performance.
Tesla has a history of fake self-driving videos [3], so this has to be looked at suspiciously.
[1] https://youtu.be/PXy90bAPwn8?t=516
[2] https://youtu.be/PXy90bAPwn8?t=776
[3] https://www.cbsnews.com/news/tesla-autopilot-staged-engineer...
SF isn't easy but Boston is next level. And it snows there. And can get heavy rain.
We don't have data on 12 yet, but it seems exceedingly unlikely they've been able to close a 3 order of magnitude gap in one release to match human ability.
I say this as an owner of FSD. It has moments where it does amazing things and moments that make me shake my head. I like it for highways, but driving in a city is too stressful to be more than a toy right now.