Why robots fail when things are both hidden and unpredictable?

Autonomous systems fail catastrophically when they face two coupled uncertainties simultaneously: they can't see their full state and don't understand how the system behaves. Researchers show this synergistic failure (77% degradation) vastly exceeds the sum of individual problems. They propose treating safety as an information problem: a metric quantifying state-dynamics coupling, active learning policies that probe dynamics, and adaptive constraints that tighten when uncertainty couples. The shift from passive robustness to strategic perception offers a path for safety-critical systems to recognize and resolve their own ignorance.