The big leap in deep neural nets, gave us the illusion that we will solve many of our problems overnight. The fact is that we probably need at least one more decade to get things right
LIDAR and geometry give you a 3D map of the environment, at least the part of it in range of the sensors. Anywhere the vehicle is going to drive should be verified as flat. That's the first step. Waymo does this. Tesla doesn't. Whether Uber does will come out when the NTSB dissects their vehicle that ran over a pedestrian.
Those top-mounted LIDAR units aren't primarily looking for obstacles. They're looking for flat road. A pothole, a cliff, a car, or a pedestrian does not look like flat road. The software doesn't have to know why it's not flat to not go there.
Trying to classify stuff that isn't flat road is mostly for dealing with moving objects. "What is that?" and "What will it do next" isn't a geometry problem. That's where neural nets come in. An acceptable result is "unidentified", which feeds back into the planner as "stay away from there".
Some people have gotten the idea that they just hook up a camera to a machine learning system, train it from a dashcam, and they have a quick path to self-driving. That will work most of the time. Until it kills someone.
Tesla claims that LIDAR isn't neccessary but yet they constantly hit barriers and stopped trucks. If they can't even get the basics right how can Musk be sure that his strategy is the right one?
The strategy is right one as long as it sells cars.
Musk is a salesman in engineer's clothing.
Tesla claims that cameras are enough and they can put it into mass-produced cars without changing the cost significantly.
Pretty large gap, don't you think?