Matt Novak of Gizmodo posted an article on April 30, 2021 with the title “Elon Musk Shares Painfully Obvious Idea About the Difficulty of Self-Driving Cars,” available at Elon Musk Shares Painfully Obvious Idea About the Difficulty of Self-Driving Cars (gizmodo.com) 
Novak quotes an April 29, 2021 Tweet by Elon Musk, as follows:
“A major part of real-world AI has to be solved to make unsupervised generalized full self-driving work, as the entire road system is designed for biological neural nets with optical images.”
Novak points out that Musk is referring to human brains and human eyes. I’m surprised that they didn’t go into that “carbon life forms” stuff.
Be that as it may, not only has it been painfully obvious for some time that the roadway infrastructure is sadly deficient in being able to support autonomous vehicles, but yours truly and others have been writing about this for years. It is indeed strange that this concept appears to have just occurred to Elon Musk, a pioneer in artificial intelligence and autonomous systems.
Here are some of my articles on the topic:
“Integrating In-Vehicle, Vehicle-to-Vehicle, and Intelligent Roadway Systems,” International Journal of Design & Nature and Ecodynamics: Special issue on “Complex Systems,” January 2018.
“Cybersecurity and Privacy Issues when Applying Railway Technologies to Intelligent Roadway Systems,” Invited Paper, Proceedings of the 2018 IEEE LISAT Conference, Farmingdale, NY, May 2018.
“Cybersecurity in the Age of Autonomous Vehicles, Intelligent Traffic Controls and Pervasive Transportation Networks,” Proceedings of the 2017 IEEE LISAT Conference, Farmingdale, NY, May 2017.
You can obtain the first article free of charge online and the other two from IEEE Xplore with the appropriate subscription.
My view on this matter is that full autonomy may not be attainable (ever?) and arguably is the wrong way to look at this whole subject. Some of my recent research has led me to realize that AI systems may never be able to achieve wisdom, understanding, emotions and true empathy and that it is pointless trying to make them do so. Sure, they can appear to be sympathetic and express concern, for example, but they don’t and won’t actually feel such emotions. The many discussions of self-driving cars, and other AI systems, about ethics and determining the right thing to do if, say. there is choice between killing the driver or a pedestrian, serve to show that such decisions are subjective and are programmed into systems based somewhat upon the views, biases and prejudices of the system designers and the data that are available to, and selected for, machine-learning.
Why don’t we just design systems utilizing the expanded capabilities and advantages of computer systems? Humans cannot detect ultrasonic sound or ultraviolet light natively. We do not have X-ray vision; nor are we able to see in the dark or ahead of vehicles in front of us; and we cannot react to certain hazards in time, or recall precisely what happened following an incident. Machines can. Humans are easily distracted, might doze off, or give over control to the vehicle when they should not. Machines aren’t and don’t.
The Biden Administration’s proposed infrastructure initiative would be a great opportunity to design smart roadways and embed the necessary sensors and beacons into them to facilitate self-driving technologies. Such an endeavor is not trivial. It involves major roadway design, prototyping and testing processes. Interestingly, Toyota is independently building a smart city to test out such ideas, as described in Faisal Khan’s April 20, 2021 article “Toyota’s ‘Woven City,’” available at Toyota’s “Woven City” envisions a fully autonomous future | by Faisal Khan | Open Source X | Apr, 2021 | Medium  Such distributed intelligent technology is key to success in a fully-autonomous world.
As an aside, I happen to think that creating a nationwide electric-vehicle charging grid, as proposed in the infrastructure plan, is a mistake. Since the electricity grids of the U.S, and other countries are so much at risk of being taken out in part by severe weather, wildfires, floods, etc. or entirely by a cyberattack by an adversary, it makes more sense to focus on microgrids. Or, even better, why not use battery-swapping as described in Roberto Baldwin’s Car and Driver article of August 21, 2020 with the title “China’s Nio Lets EV Drivers Swap Batteries in 5 Minutes, Hit the Road,” available at China’s Nio Lets EV Drivers Swap Battery Quick and Hit the Road (caranddriver.com)  ? Furthermore, Baldwin quotes Mark Gillies of Volkswagen as saying: “Our data indicates that only 3 to 5 percent of all EV drivers use fast charging as an option …” with most charging at home. If that will be the case, a nationwide electric-vehicle charging grid makes even less sense.
We might have an opportunity to really modernize transportation systems and infrastructure. Let’s ignore the smoke-and-mirrors nonsense that seems to pervade this area and design and build systems that utilize technology to its full advantage and provide the kind of transportation network that can work—and that involves smart roadways.