Automated Driving Safety Framework

• Safety Integrity Framework for Automated Driving is an integrated structure combining functional safety, SOTIF, and AI-specific standards to manage and mitigate system risks.
• It outlines a lifecycle from hazard identification and risk assessment to mitigation design and continuous monitoring using formal verification and traceability.
• The framework employs detailed quantitative risk metrics, redundancy patterns, and an AI Data Flywheel to maintain safe operations amid open-world uncertainties.

A Safety Integrity Framework for Automated Driving defines the architecture, methodologies, artifact flows, and quantitative metrics required to establish and maintain a demonstrably acceptable level of risk for autonomous road vehicles throughout their development and operational lifecycle. Such frameworks must unify functional safety (ISO 26262), Safety of the Intended Functionality (ISO 21448), and dataset/AI-specific standards (ISO/PAS 8800), while providing continuous assurance amid the stochastic, open-world uncertainties inherent to automated driving systems (ADS) (Patel et al., 4 Mar 2025, Werling et al., 26 Mar 2025, Abbaspour et al., 11 Nov 2025).

1. Foundational Standards and Conceptual Underpinnings

Safety Integrity for automated driving is governed by the integration and extension of three principal standards:

• ISO 26262: Focuses on malfunctioning behavior in E/E systems, using the Automotive Safety Integrity Level (ASIL) scheme to derive requirements from Severity, Exposure, and Controllability ratings. PFH (Probability of Dangerous Failure per Hour) must remain below graded thresholds (e.g., ASIL D: PFH < $10^{-8}$ h$^{-1}$) (Vyas et al., 2024).
• ISO 21448 (SOTIF): Addresses non-malfunctioning behavior, targeting unreasonable risk from performance limitations, unknown scenarios, specification insufficiency, and foreseeable misuse. SOTIF partitions scenario space into four classes and aims to reduce risk in all hazardous areas by improving both the specification and performance of the ADS (Patel et al., 4 Mar 2025).
• ISO/PAS 8800: Specifies data lifecycle and AI-specific assurance processes, covering dataset safety requirements, gap analysis, annotation quality, versioning, and continuous verification/validation (Abbaspour et al., 11 Nov 2025, Abbaspour et al., 5 Feb 2026).

Central tenets include quantitative risk representation, explicit risk acceptance criteria, embedded redundancy, continuous monitoring, and unified traceability linking safety claims to technical evidence (V-model, scenario catalog, runtime evidence) (Salem et al., 2023, Werling et al., 26 Mar 2025).

2. Safety Integrity Lifecycle Phases and Core Processes

The safety integrity process for automated driving proceeds through the following lifecycle phases, each mapped to a spectrum of standards and utilizing both classical and AI-centric methods (Vyas et al., 2024, Patel et al., 4 Mar 2025, Werling et al., 26 Mar 2025, Abbaspour et al., 11 Nov 2025):

Phase	Activities	Dominant Methods/Standards
Hazard Identification	Scenario/cause cataloging, misuse, ODD gap analysis	STPA, FTA, HAZOP, Rulebooks, ODD-monitoring
Risk Assessment	Probability/severity quantification, acceptance match	Bayesian networks, QRN, FMEA/PFMEA, SOTIF QRN
Mitigation Design	Redundancy, fallback, AI/ML risk control, Safety Shell	RNFTC, Safety Shell, AMLAS, architecture patterns
Verification & Validation	Scenario-based testing, simulation, fault injection	SIL/HIL/VIL, genetic algorithms, digital twin, boundary analysis
Continuous Monitoring	Risk/dataset drift monitoring, field-data feedback	SOTIF entropy, operational monitoring, retraining loop

In data-centric processes, the AI Data Flywheel governs closed-loop dataset iteration to ensure coverage of the Operational Design Domain (ODD), robust annotation, and validation of rare corner cases, tightly coupled to safety requirements (Abbaspour et al., 11 Nov 2025).

3. Quantitative Risk Metrics, Scenario Coverage, and Uncertainty Management

Central to the framework is explicit risk quantification, unifying aleatoric (scenario, hardware, situational) and epistemic (model, knowledge) uncertainties:

• Risk Metric: $R = \sum_i P_i \times S_i$ for each hazard $i$, where $P_i$ is occurrence probability and $S_i$ is severity (Salem et al., 2023).
• Residual Risk: $R_{\text{residual}} = \sum_i P_i' S_i$, with post-mitigation probabilities $P_i' = P_i^0 \times (1-I_{m,i})$ based on safety integrity $I_{m,i}$ of each measure (Salem et al., 2023).
• Acceptance Criterion: $R_{\text{residual}} \leq R_{\text{tolerable}}$, e.g. matching observed risk of human drivers or a regulatory bound (e.g. $^{-1}$0 fatalities/hour at crosswalks) (Salem et al., 2023, Werling et al., 26 Mar 2025).
• Scenario Coverage: Formalized in both sample-based ($^{-1}$1) and formal ($^{-1}$2) terms, leveraging scenario volume metrics and logical/temporal safety contracts (Zhao et al., 2022).

Probabilistic simulation, Bayesian updating, and sensitivity analysis are employed to propagate all uncertainties through the hazard–mitigation–outcome chain, allocate risk-reduction efforts, and drive iterative architectural improvement (Werling et al., 26 Mar 2025).

4. Safety Architecture Patterns: Redundancy, Degradation, and Online Risk Arbitration

Frameworks at the system-architecture level realize safety integrity via:

• Layered Redundancy: Multi-channel architectures (e.g. Safety Shell, DSM) use diverse perception, planning, and fallback mechanisms to ensure fail-operational performance under dual-point faults, minimizing the dependence on high-ASIL single points (Hanselaar et al., 2023, Fu et al., 2020).
• Degradation Policy: State machines step through operational, degraded detour/rescue, and fail-safe emergency modes in response to detected faults or ODD boundary violations (Fu et al., 2020). Each step is formally verified using modal μ-calculus to ensure that safe-stop is reachable under all modeled multi-point faults.
• Online Arbitration: At run time, cross-checked world models and motion plans are dynamically selected based on predicted risk (e.g., maintaining $^{-1}$3), ensuring both ongoing availability and safety by switching to the channel with maximal anticipated margin, or triggering an emergency fallback if immediate danger is identified. Risk thresholds explicitly map to regulatory SOTIF/FuSa concepts (Hanselaar et al., 2023).
• Scenario-Based Assessment: Systematic scenario extraction from real data, importance sampling, and simulation produce quantitative KPI measures (e.g., collision probability $^{-1}$4, time-to-collision CDFs), supporting regulatory approval and field monitoring (Ploeg et al., 2021).

5. Formal Verification, Requirement Traceability, and Continuous Assurance

Every substantive framework embeds formal verification mechanisms and traceability artifacts:

• Formal Methods: End-to-end functional safety is supported by program logic (dFHL, STL contracts), reachability analysis (maxFRS/BRS), model checking, and Hoare annotations, enabling mathematically rigorous safety property proofs even in intersection and multi-agent scenarios (Haydon et al., 2023, Zhao et al., 2022).
• Requirement Management: All safety requirements (behavioral, technical, dataset) are maintained with bi-directional traceability linking hazard/risk analysis, test results, artifact lineage, and operational feedback. Example: dataset requirements DS-RQ1/2 (Abbaspour et al., 11 Nov 2025) are mapped to top-level AI safety goals and SOTIF/ODD definitions.
• Continuous Feedback: Continuous operational and dataset monitoring, field-data integration, and ongoing risk evaluation (e.g., SOTIF entropy) provide closed-loop assurance, with enforced re-certification on model or data drift, and explicit connection to change management (Patel et al., 4 Mar 2025, Abbaspour et al., 5 Feb 2026).

6. Advanced AI and Dataset Integrity Measures

Given the centrality of AI/ML in perception and decision-making, comprehensive dataset and model integrity control is required:

• AI Data Flywheel: Governs the lifecycle from data collection, VLM-based scenario selection, annotation, QC, retraining, and deployment, with strict version control and traceability to all requirements and safety analyses (Abbaspour et al., 11 Nov 2025).
• Hazard and Risk Analysis: Extended with HAZOP, FMEA/PFMEA, STPA at all lifecycle stages; e.g., detecting corner-case data insufficiency or annotation error propagating to residual model risk (Abbaspour et al., 11 Nov 2025, Kou et al., 21 May 2025).
• Dataset V&V: Employs equivalence-class, boundary analysis, error guessing, and coverage/interference metrics (mAP, IoU, explicit scenario coverage ratios). Distributional drift is monitored (e.g., $^{-1}$5), and test cases are stratified across all relevant ODD conditions (Abbaspour et al., 11 Nov 2025).
• AI-Specific Safety: AI/ML requirements (robustness, fairness, reliability) are decomposed into dataset design, model V&V, and runtime monitoring steps, each aligned to ISO/PAS 8800 and integrated into the system V-model (Abbaspour et al., 5 Feb 2026).

7. Ongoing Challenges and Synthesis

Despite the rigor of current frameworks, ongoing challenges include quantifying representativeness and completeness, integrating dataset assurance directly into top-level safety cases (ISO 26262/21448), and securing data/model pipelines against adversarial attacks (Abbaspour et al., 11 Nov 2025, Abbaspour et al., 5 Feb 2026). Extensions under active exploration include probabilistic and data-driven reachability, deep uncertainty quantification, and AI-driven scenario synthesis for scalable validation (Zhao et al., 2022, Patel et al., 4 Mar 2025).

This synthesis demonstrates that state-of-the-art safety integrity frameworks for automated driving are fundamentally cross-disciplinary, embracing explicit risk metrics, scenario-centric and formal methods, dataset-centric AI assurances, continuous lifecycle monitoring, and robust organizational traceability. Only such holistic, quantitatively grounded approaches can provide credible safety assurances commensurate with the complexity and societal criticality of full-stack autonomous driving (Werling et al., 26 Mar 2025, Hanselaar et al., 2023, Patel et al., 4 Mar 2025, Abbaspour et al., 11 Nov 2025, Abbaspour et al., 5 Feb 2026).