How are you stumbling and re-stumbling on most edge cases in the real-world? The comparison to aircraft development is a bit misleading. The aircraft folks use simulation for most development and testing. And they do not use the public domain for most of that nor public test subjects.Those aircraft simulators and simulations are so precise now they build them before most of the aircraft test versions are built. An analogy is some of the vehicle models in this industry. Unfortunately the associated simulation technology beyond that in this industry is inadequate. the technology this industry uses cannot create an all model type physics based digital twin. And run that in proper real-time fully loaded and with complex scnerios. DoD simulation tech can. Given all of this the real-world is certainly a reference but you cannot use it to find, create, design for or test most edge cases. (I would be glad to demonstrate the DoD simulation technology I referred to.)
With regard to machine learning. Deep learning is a significant issue. It requires far too much time, cost and processing power. And it is too easily confused by patterns, shadows etc. The industry should switch to dynamic sensing and avoidance with targeted seep learning
An HD Map to Avoid the Crash of the Autonomous Vehicle Industry —
Proposal for Successfully Creating an Autonomous Ground or Air Vehicle
