They need sensor fusion. The system needs to make maximum use of all the information available to it: Where is the road striping? Where are the other cars going? Where are the road signs and signals? (If there's one in your path, you certainly shouldn't drive into it!) Are there camera-visible obstructions? What were the interpretations and actions of previous Tesla trips along the same route?
In these problem cases, all data except the left and right lane striping seems to be completely ignored. There was even more information at the fatal offramp location (cross-striping over the lane separation zone), which the vehicle drove straight over. The system is not making maximum use of the information available to it, in fact it is using hardly any of it at all, and fixating on what it thinks is a single most salient piece of data.
Sensor fusion algorithms tend to behave the opposite way-- each additional piece of data informs the interpretation of all the other data. You can have very poor-quality data, but if it is even moderately over-constrained, your state estimate can be very good in spite of it. I think it would be completely reasonable to have a neural net in the loop of a sensor fusion algorithm, with fusion constraints informing the NN's interpretation, and the NN's estimates feeding back into the fusion algorithm as uncertain data.
IMO Tesla will do at least one of:
* Very expensively retract their promise of full self driving for delivered vehicles
* Completely overhaul/redesign their driving software and start again nearly from scratch
* Get into a regulatory/legal tangle with the NTSA/courts/DOJ over all the dead people their system is making.