Safer Than Perception: Increasing Resilience of Automated Vehicles Against Misperception

Fränzle, Martin; Hein, Andreas · 2025 · Crossref

DOI: 10.1007/978-3-031-73741-1_25

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety gap between the high error rates of machine-learning-based perception systems in autonomous vehicles (AVs) and the stringent societal safety requirements for automated driving. The authors argue that while individual perception components (such as object detection and classification) have error rates orders of magnitude higher than acceptable safety thresholds, the overall system can be made "safer than perception." This is achieved by designing guard conditions for safety-critical maneuvers that are resilient to misperception, ensuring that the frequency of erroneous maneuver activation is significantly lower than the frequency of individual perceptual errors. The study proposes a methodology for constructively generating reformulations of Boolean guard conditions that are more robust to uncertainty. The approach relies on the concept of "world equivalence" (W-equivalence), where two logical formulas evaluate identically with respect to ground truth under specific environmental invariants but may have vastly different error rates when evaluated against noisy percepts. The authors demonstrate that by exploiting these invariants, one can rewrite guard conditions to minimize false positives (safety risk) while maximizing true positives (system availability). The method formulates this optimization as a constrained problem, utilizing integer-linear programming (ILP) to synthesize optimal guard condition representations that satisfy a given societal risk threshold. Key findings include a rigorous mathematical proof that complex Boolean combinations of atomic percepts can exhibit significantly lower misevaluation rates than their constituent parts. Using a simplified example involving flat obstacles, the authors show that rewriting a guard condition to require a higher threshold of detected grid cells (while remaining W-equivalent to the original condition) can reduce the false-positive rate by approximately 16-fold compared to the original formulation. For instance, with atomic classifier false-positive rates of 0.2, the optimized guard condition achieved a false-positive rate of roughly 0.061, well below the atomic error rate, while maintaining a true-positive rate of 83.4%. The paper further presents an algorithmic approach to automatically synthesize these optimized conditions by treating impossible ground-truth states as "don't-care" entries in truth tables, allowing for the strategic assignment of truth values to minimize risk without compromising logical correctness. The significance of this work lies in providing a constructive, formal method to bridge the reliability gap in autonomous driving systems. It challenges the prevailing focus on isolated optimization of computer vision components, suggesting instead that system-level safety can be enhanced through logical structuring of decision-making guards. By enabling the automatic synthesis of resilient guard conditions, the methodology offers a practical pathway to ensure that automated vehicles meet strict safety standards despite the inherent uncertainties of current perception technologies. This approach complements existing analytical frameworks by providing a tool for actively optimizing system resilience rather than merely analyzing it.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness