I'd offer my bets (you can guess from earlier posts in this thread), but I think the answer to that is largely going to be both subjective and based on very sparse information about proprietary implementation details. The only data we're really going to see as the general public is more-or-less on-the-road performance metrics, like collisions, close calls, misbehaviors, hard braking incidents, etc -- what's obligatorily being reported to NHTSA, state licensing agencies, and overseas counterpartsI was reading a reddit post or something where someone said the old Uber ATG systems were a pile of nvidia gtx 1080s in the trunk.
From what I understand today, the lidar is cheap and easy enough, and now the processing and computing hardware is also commodified enough (between nvidia and tesla and whoever) so now the secret sauce is all in the software. Who has the best lidar -> driving software stack?
If @demultiplexer gets his wish, maybe we'll see data about performance in various standardized testing scenarios, too. I'm skeptical that'll ever happen, because the companies involved each see advantage to improving on-the-road performance from keeping their narrower functional testing methodology and data proprietary. If the on-the-road performance is good enough, regulators would be hard-pressed to justify a demand for that data anyway. IIRC, Waymo and Mercedes have each published some details in this vein, though.