Initial stabs at self-driving were a tangle of neural networks and hard-coded rules, requiring vast quantities of real-world data to finesse. Key to the process was not just using regular maps, but minutely surveying areas of operation, which requires constant updates.

The appeal of Tesla’s approach is that it avoids this step, using camera feedback to make on-the-fly decisions.

The likes of UK-based Wayve and China’s Pony AI (PONY.O), opens new tab, meanwhile, are using generative AI to create virtual training scenarios to prepare their models for the statistically unlikely but potentially ride-ending “corner cases” that human drivers can readily handle.

This part of the article was weird to me. What it really should explain is that Tesla's camera approach is much harder technologically and probably requires a lot more data to train, but they don't have to make detailed 3D maps like LIDAR based approaches do so they aren't going to be restricted to operating in certain cities. And it won't take a ton of work to expand to new locations. Instead it half communicates that idea in a sprinkle of other half communicated ideas that would probably end up giving the reader the wrong impression about the differences between Tesla and other AV companies.

When I worked at an AV company 2 years ago, most companies were definitely still using a mix of neural networks and hard coded rules. End-to-end neural networks were kinda in the works but not quite reliable yet. I don't know if they've been achieved since then but regardless that has nothing to do with why some companies need precise mapping. The maps provide the LIDAR a point of reference so the car can more easily determine exactly where it is in its surroundings, and so that it knows which objects are things that are part of the environment like walls, trees, etc, and which are things that it should expect to move. Tesla doesn't need that because they have no LIDAR.

A camera only approach like Tesla's needs a shit load of data, likely a lot more than LIDAR based approaches. If their claims of using a single massive end to end neural network are true, then that needs even more data. Essentially all of the information that LIDARs give you "for free" now needs to be replicated by a neural network looking at only cameras so that network probably needs more data. One of Tesla's major advantages and why they can be even remotely competitive with such a handicapped set of sensors is probably because their existing customers are more than happy to help them do a public beta test on millions of vehicles out on the road and collect data.

Tesla almost certainly does simulated testing too, like every other AV company. And they probably use gen AI in at least part of their simulation pipeline.