Unless the power is out

https://abc7news.com/post/san-francisco-leaders-press-waymo-...

Kind of unrelated. That issue was due to a misguided effort to be cautious by having vehicles requesting human-review when they didn't really need it. Waymo fixed the issue by allowing the vehicles to operate in their normal, independent, mode.

part of the problem is that SFs traffic lights just turn off in a power outage, rather than flashing red battery power as I have seen in many other jurisdictions

This seems tautological, but in practice, you might expect to see different results.

Engineering hours are finite, so if they're spread across interpreting signals from two different sources, they might not go deep enough to make either one as good as it could be.

Having your engineering resources more focused on a particular approach might actually yield better results.

I say this as someone who's dealing with LiDAR + vision vs pure vision in a different domain, and at this point, I actually think our pure vision systems are better.

More often than not, constraints refine and focus a project, rather than restricting it. It’s best to start work with as few variables as possible, and only add new ones when absolutely necessary; You make a lot more progress that way.

For very complex things like AVs, it is critically important to keep the number of such variables down, since each acts on complexity & workload not as an addition but more like a quadratic, or worse—combinatorial explosion.

What if they oppose each other? Which one do you trust?

Google ”Sensor fusion”.

Also google "sensitivity-specificity tradeoff." Initiating emergency braking when at least one sensor detects a pedestrian has the highest sensitivity, but the worst specificity. The real trolley problem of self-driving: how many spilled drinks and read-end collisions society is willing to tolerate to save one life.