"I have detailed knowledge of GM/Cruise's software. It is of unusually low engineering quality, and should not be viewed as comparable to Waymo's high-quality, carefully-written software.
GM/Cruise's system was written rapidly, under extreme time-pressure, by junior engineers. It is riddled with shortcuts and unsound partial solutions. The system exists for the purpose of giving a good demo (to SoftBank, for example).
GM/Cruise's system is based on ROS (
en.wikipedia.org/...). ROS is a framework that is appropriate for school projects but cannot realistically be used in cars or aircraft or other latency-critical applications. ROS allows fast, easy development (especially for junior developers) but the result will never really work, in the sense that the cars cannot be expected to operate without a Safety Driver present.
Look past GM's advertising campaign (the "news articles" they buy). Realize that GM/Cruise has serious long-term problems coming down the pipe.
The GM/Cruise cars follow hand-drawn paths through the city. There is a big group of people sketching out routes for the cars to follow, like trains on tracks. Adjustments are made according to a corpus of hand-coded rules. The system is dumb (does not generalize, in the AI sense) but can handle the normal situations it has been coded to handle. You can rack up a lot of miles like this, following tracks over and over again, with code to handle the common cases.
When something weird happens, the Safety Driver (a human) takes over to prevent disaster. For the self-driving car to make sense financially, as a business, you can't have a Safety Driver babysitting the car. I mean, he may as well just be driving it. GM/Cruise's software works most of the time, but it's not really reliable because it can't properly handle the "long tail"-- the numerous different events that are each quite unlikely but probable in aggregate.
Waymo's software is truly smart. GM/Cruise's software is dumb and sloppy and fake, and ultimately a bad investment.
