Most current driving devices are nevertheless skilled and evaluated on naturalistic eventualities gathered from each day life or heuristically created adversarial kinds. On the other hand, the huge populace of cars in normal sales opportunities to an exceptionally low collision fee, indicating that the security-significant situations are scarce in the collected real-globe data. Consequently, methods to artificially generate security-vital scenarios develop into important to evaluate the risk and lessen the value. In this converse, I will initially supply a in depth taxonomy of present algorithms by dividing them into a few groups: data-pushed technology, adversarial era, and expertise-based generation. I will then introduce a number of certain algorithms of my former do the job. Lastly, I will prolong the discussion to five principal worries of current performs — fidelity, performance, variety, transferability, controllability — and exploration chances lighted up by these problems.
