Autonomous systems are designed for repetition. They are good in the situations where patterns can be memorized, charted, and anticipated with a high level of certainty. However, real-world driving is filled with edge cases, which do not scale well to datasets.
Even a sophisticated system can be thrown off by a plastic bag floating along a highway or an unmarked detour in a construction area. The anomalies are intuitively processed by human drivers.
They do not need prior exposure to such scenarios since they reason by analogy rather than data. This generality ability with respect to known inputs is a structural benefit. [Read more…] about Where Drivers Still Beat Autonomous Systems, and Why it Matters