Anticipatory Thinking: A Testing and Representation Challenge for Self-Driving Cars

Abstract

Ensuring autonomous systems can reason and anticipate unknown (erroneous) futures is important for safety, trust, and reliability. Self-driving is an important domain because of the difficulty in testing: not all erroneous scenarios cannot be covered in training nor simulation experiments. I propose a series of tests for autonomous vehicles that require anticipatory thinking: the deliberate and divergent exploration of relevant possible futures. These stress tests include abstract thinking like theory of mind or self-introspection instead of memorization. Developing this capability, results in adaptive, autonomous vehicles that can reason and address the ever-growing, long tail of errors.