SF is unpleasant to drive as a human, but slow, dense traffic seems like a near ideal scenario for autonomous. SF needs lots of social calculations if you don't want to get honked at, which is mentally taxing, but they're far from necessary for safety.
Even fairly simple autonomous tech will have better peripheral vision at near-to-mid range than one human can manage, so for all those bikers, crazy walkers, and chaotic 15mph cars you shouldn't hit, it stands a pretty good chance of being better. And when it's not, come to a stop and you're fine (barring some honks) - few are moving fast enough to hit you dangerously hard in those human-complex areas, and you don't need to stop instantly, just fast enough.
---
Honestly, I'd put SF at dramatically easier than either residential or highway roads. Residential (and adjacent) has fast-moving cars ignoring signs with obstructed vision, and inattentive humans at relatively high speed (bikers used to low traffic and swerving, kids and animals running literally through bushes adjacent to roads, general lack of care around vision-blockers like fences due to perceived low risk, etc). Who's-at-fault doesn't matter - in car vs human, humans lose, and people rightfully get upset.
Highways also seem harder, if highly specialized: accurate decisions 100+ feet in advance are absolutely critical due to the speeds involved, computer vision at that range has fairly low detail compared to humans, and lidar is practically braille for "car". Radar has trouble distinguishing stopped cars from the road because neither are moving. Ultrasonics as an ultimate backup really only work up to around 10m (and that's about the distance to stop a car at 25MPH, which you'll regularly encounter in dense city traffic).
I'll also point out that more people have died due to Tesla's autopilot on highways than Waymo, Uber, Cruise, heck all self-driving companies I'm aware of at any size combined. They're riding all the terminology lines they can to get away with it, and they may very well have an order of magnitude or two more miles, but I believe the point still stands - highways are hard.
Even fairly simple autonomous tech will have better peripheral vision at near-to-mid range than one human can manage, so for all those bikers, crazy walkers, and chaotic 15mph cars you shouldn't hit, it stands a pretty good chance of being better. And when it's not, come to a stop and you're fine (barring some honks) - few are moving fast enough to hit you dangerously hard in those human-complex areas, and you don't need to stop instantly, just fast enough.
---
Honestly, I'd put SF at dramatically easier than either residential or highway roads. Residential (and adjacent) has fast-moving cars ignoring signs with obstructed vision, and inattentive humans at relatively high speed (bikers used to low traffic and swerving, kids and animals running literally through bushes adjacent to roads, general lack of care around vision-blockers like fences due to perceived low risk, etc). Who's-at-fault doesn't matter - in car vs human, humans lose, and people rightfully get upset.
Highways also seem harder, if highly specialized: accurate decisions 100+ feet in advance are absolutely critical due to the speeds involved, computer vision at that range has fairly low detail compared to humans, and lidar is practically braille for "car". Radar has trouble distinguishing stopped cars from the road because neither are moving. Ultrasonics as an ultimate backup really only work up to around 10m (and that's about the distance to stop a car at 25MPH, which you'll regularly encounter in dense city traffic).
I'll also point out that more people have died due to Tesla's autopilot on highways than Waymo, Uber, Cruise, heck all self-driving companies I'm aware of at any size combined. They're riding all the terminology lines they can to get away with it, and they may very well have an order of magnitude or two more miles, but I believe the point still stands - highways are hard.