Safety of the Intended Functionality for Automated Driving

Autonomous driving failures often occur without hardware or software defects, caused instead by sensor blindness, heavy rain, or corner-case edge scenarios. SDVSafety investigates ISO 21448 SOTIF methodologies, runtime monitors, and safe stop maneuvers.

Resolving Perception Uncertainties and Performance Limitations

ISO 26262 addresses internal system faults; ISO 21448 focuses on hazards caused by functional limitations and environmental complexities. Engineering automated driving requires continuous risk estimation and deterministic validation in unknown operational spaces.

Core Engineering Areas

• ISO 21448 SOTIF: Safety of the Intended Functionality — ISO 26262 functional safety assumes that systems operate safely if components do not fail mechanically or electrically. However, an autonomous vehicle camera blinded by direct low-angle sunset, or a radar confused by wet guardrail reflections, has zero component hardware faults yet produces life-threatening decisions. ISO 21448 (SOTIF) systematically minimizes the 'Unknown-Unsafe' operational scenarios.
• Operational Design Domain (ODD) Enforcement and Minimal Risk Maneuvers — No autonomous driving system is capable of driving everywhere in every condition. The Operational Design Domain (ODD) formally specifies the environmental, geographical, and speed constraints under which an automated feature is certified. When heavy snowfall or road construction breaks the ODD boundary, the vehicle must execute an automated Minimal Risk Maneuver (MRM), steering onto the highway shoulder without human intervention.
• Safety Monitors for Neural Networks: Runtime Out-of-Distribution (OOD) Detection — Deep neural networks used in autonomous vehicle perception are fundamentally uninterpretable black boxes that often predict incorrect classes with catastrophic overconfidence. Runtime Out-of-Distribution (OOD) safety monitors evaluate intermediate feature embeddings, Mahalanobis distance metrics, and energy-based scores to detect when visual input diverges from training data, triggering safety envelopes before an accident occurs.

Technical Articles

• SOTIF vs Functional Safety: Why ISO 26262 Is Insufficient for Autonomous Driving
• Minimal Risk Maneuvers in L3 Autonomous Driving: Pulling Over Safely on the Highway
• Runtime Out-of-Distribution (OOD) Detection for Vision Transformers in Autonomous Vehicles

Related Cluster Architecture Specifications

• AutomotiveSDV — Full SDV Stack & Reference Architecture
• A2HCard — Agent-to-Human (A2H) Interface Cards & In-Cabin Autonomous Interaction Manifests
• A2HumanCard — Standardized In-Vehicle AI Agent Capability Discovery, Trust & Authorization Cards
• AIDrivenSynBio — AI-Driven Synthetic Biology for Automotive In-Cabin Air Purification, Bio-Engineered Pathogen Scavengers & Molecular Biosensors