My understanding is that image processing and recognition is the big Achilles heel for high power computer systemsat the moment.
Developing say a FBW architecture, even the learning/damage compensating kind is easier as the inputs and outputs are both very easily specified and provided. However, the ability to use visual/IR imaging and then to make sense of that picture and act appropriately is very challenging at the moment.
Self-driving cars manage because road signs and cars are simple, predictable shapes and easy for an image recognition program to "learn" - an autonomous or optionally crewed aircraft out there will not have at luxury, but be able to visually interpret anything the world can throw at it.
This is one task the human brain is bloody brilliant at compared to computers. The oft-quoted example is some,body throwing a tennis ball at you - a human has seen, recognised and caught it before a modern computer has identified it and the way the brain recognises objects is still somewhat of a mystery so we don't really know how we can train computers to do it efficiently.
Autonomous flights between Heathrow and Frankfurt? Maybe. Autonomous flights into any Greek island of your choice with nothing but a ****e VOR or a visual? That is the real challenge for automated air transport - to match the flexibility of the current system without massive infrastructure expenditure.