SYSTEMATIC, DATA-DRIVEN APPROACHES FOR AUTONOMY TESTING AND CERTIFICATION - (PPT)

摘要

"Shift from the simple, binary view of standards to a staged, dynamic evaluation resembling the approval process for drugs" (which also happen to be inscrutable black boxes). Design controlled experiments on the deep learner to: (1) Generate data about the actual safe operational domain and; (2) Find suitable spaces in which we can compare the similarity between given operating environments. Successful "early trials" permit ever more expanded operational domains. This isn't just testing to rack up miles, these would be scientific and reviewable experimental programs.