LIDAR is not necessarily categorically mandatory, but shadow braking and other autopilot failure modes demonstrate that it or something with its capabilities are probably necessary right now to enable today's "AI" to safely drive a car. My point about active sensing being able to decide edge cases explains why this is likely the case.

IMHO doing without precise active sensing would require a breakthrough in AI that hasn't happened yet.

On the plus side the fact that Tesla's autopilot is able to work as well as it does without lidar might indicte that with a lidar system it might be the best autopilot on the market. Who knows.

Seems weird that the Tesla can't infer that there is no wall there from GPS data.

What if a wall is there tomorrow?

They do over time. The vehicle “learns” from no impact at the overpass and uploads that to the mothership for fleet-wide distribution.

Obviously LIDAR is not mandatory. We have a billion proof of concept controllers that only use binocular image data to navigate just fine. So one day ostensibly that could be enough. But we need a computer as good as the human brain to make it work, and that's not happening anytime soon. So until then, dropping LIDAR from the sensor suite is really going to limit your performance.