Abstract
<title>Abstract</title> <p>Artificial intelligence and machine-learning technologies have become fundamental enablers of perception, prediction, trajectory planning, decision-making, and motion-control functions in advanced driver-assistance systems and automated driving systems. Nevertheless, the integration of learning-enabled components into safety-critical automotive architectures remains constrained by the limited compatibility between conventional deterministic software-assurance methods and the data-dependent, probabilistic, and context-sensitive behavior of machine-learning models. Although the AUTomotive Open System ARchitecture (AUTOSAR) Adaptive Platform provides a service-oriented execution environment suitable for high-performance computing, sensor fusion, and autonomous-driving applications, AUTOSAR does not independently constitute a complete assurance framework for artificial intelligence. This paper proposes an AUTOSAR-AI Safety Assurance Framework (AASAF) for integrating AI-enabled autonomous-driving functions into automotive safety lifecycles. The framework combines AUTOSAR Adaptive Platform services with ISO 26262 functional safety, ISO 21448 Safety of the Intended Functionality, ISO/SAE 21434 cybersecurity engineering, ISO 24089 software-update engineering, and AI-specific risk-management and robustness guidance. A complementary AI Assurance Tier model is introduced to determine the required rigor for data governance, model verification, uncertainty estimation, operational-design-domain monitoring, runtime supervision, authority limitation, and deterministic fallback. The proposed tiers do not replace Automotive Safety Integrity Levels and are not interpreted as equivalent safety classifications. Instead, they tailor AI specific evidence and runtime controls according to the safety criticality and operational authority of the learning-enabled component. The framework introduces a safety-supervised AUTOSAR architecture incorporating an AI Assurance Manager, independent runtime monitors, a deterministic Safety Authority Manager, and a verified Minimum Risk Manoeuvre controller. A structured mapping between AI hazards, ASIL-oriented safety goals, AUTOSAR services, and verification evidence is presented. The resulting architecture supports controlled deployment of AI within autonomous-driving systems while maintaining traceability, freedom from interference, cybersecurity protection, safe degradation, and lifecycle assurance.</p>