diplomat33
Average guy who loves autonomous vehicles
Traffic laws are perfect for hard coding, though. ML is imprecise and computationally inefficient. Ideally you've have simple, drop-in traffic law modules for each location. No right on red in Manhattan (and various other places) is literally a couple lines of code, and that code will be 100% reliable. Why maintain a separate neural network for places that don't allow right on red and spend all that time gathering thousands of "don't turn on red" scenarios to train that separate neural net? And still run the risk it might screw up?
The trick is melding the hard-coded traffic law module with generic NNs that handle the bulk of the driving task. No issue for perception and not a big issue for planning, but could be tricky for prediction.
It's stuff like this that make me extremely suspicious of end-to-end claims. Including Musk's.
You raise a good point. It is why I am skeptical of E2E as well. It just seems like a very hard approach. We know road infrastructure is terrible in the US. It is not uniform. And there are big differences in traffic conditions, rules and roads. So training one single NN, with no hard coded rules, to understand the vast differences in rules, traffic and roads in the entire US, seems unlikely. I feel like Elon is attracted to the simplicity and elegance of E2E but it blinds him to the challenges of executing the approach.