The two sensors operate on different physical principles and their failure modes are nearly uncorrelated; therefore, the cross-modal discrepancy-driven approach is more robust than single-model uncertainty sampling.