There's no evidence that Cruise is in the same class as Waymo.
determining the path of travel through lanes markings that suddenly fade out while moving in 70mph on a busy freeway requires advanced AI
its pretty clear the self-driving gods are not giving these cars a grounds up understanding of physics and self-preservation. They best thing they come up with is programmed to drive X mph and to respond probabilistically to a large but finite set of training input scenarios.
Tesla is relying upon advanced AI to do things that its sensors can't do reliably. Its only Tesla who is trying to make self-driving cars off of cameras and inaccurate radar alone.
Maybe Lidar really is too expensive for mass-market adoption. But its the most obvious solution to the problem right now, instead of trying to solve harder problems (ie: camera recognition of 3d objects over trained neural nets).
Is it? Unless you are talking about NASCAR, an adversarial game where rational agents compete for road space, driving is essentially a physics problem: maintain a speed and direction that does not exceed the vehicle's ability to correct them so as to avoid any stationary or mobile obstacle.
In this case, the vehicle was accelerated to a speed (72mph) that far exceeded its safety speed against oncoming stationary barriers (0mph).
I concede that discerning what is an obstacle and what is drivable road surface is a hard computer vision problem, hence the tendency to simplify it using 3D data, radar/lidar etc. But still, not quite general AI.
- perfect cv image classification (+ low visibility )
- anticipating behavior of bikers, children, elderly, animals, near roads
- responding to human non-verbal communication
- responding appropriately to never before seen obstacles/scenarios without being a nuisance on the road or endangering other drivers
- vehicle dynamics (+ in rain, ice, snow, emergency etc)
...
I would guess to do these tasks with accuracy similar to humans approaches a problem space that's as difficult as general AI
some of these seemingly require the machine to have theory-of-mind which is general AI
The same happens with humans. Reason is extremely slow. Pilots are trained to act fast training the subconscious, not the logical mind.
Hence, the opportunity for an automated system that does not do that to be much safer, by relying on reaction times rather than strong AI.
Where a human driver would use subtle cues to anticipate a slowdown of the preceding vehicle before a turn, an automated system can get by by simply slamming the brakes in less than one millisecond when it detects braking from the other vehicle.
those types of twitchy driving mechanics aren't normal and don't share the road well with normal humans. The physics of cars and reaction times would dictate that we program a self-driving car to drive like Grandma .... always maintain safe low speeds, very long following distances and braking hard for sketchy actions by other cars or random things near the road, but we know that actually makes the road more dangerous as humans drivers will aggressively cut-off and rear end this self-driving grandma. Secondly, that type of driving creates a bad public impression of self-driving cars hurting their chances of adoption. If you read some of the earlier impressions of Google/Waymo cars its clear they went down that path initially and had to change their approach.