My current understanding is Waymo went with LIDAR + ML and Tesla is still betting on a neural net approach, but I’m looking to appreciate the differences in approach to solving hard problems with AI.
For example, their cars need not reliably detect roads or traffic signs because the database shared between all cars knows where they are. If a traffic sign is changed, and one of their cars detects it, they’ll update their database, and all other cars will know to expect to see it there (they may do a manual check afterwards, or require n > 1 cars to report the same)
This also makes it easier to detect obstacles, as the system can know what every lantern post, traffic sign, advert, etc. on the road looks like.
They very likely have on-board systems to be a bit smarter, but they could simply brake or stop the car whenever their on-board systems ‘see’ something that’s sufficiently different from what the model says they should see.
Tesla, on the other hand, wants their cars to work autonomously. In theory, you can drive one everywhere in the world, including places no Tesla has ever been to (in practice, their models may not be trained well on foreign roads)
They are both using AI.
The main question is whether self driving is possible nearly everywhere with a millisecond budget and X flops of compute. Where X is a reasonable cost for a consumer.