Standards & resources | Automotive AI Safety Forum

Content type

Topic Source language Source origin country Project region Format Publisher Record year Publication status

Research papersarXivPosted 2026.10.29EnglishPreprint

LiDAR-SHIELD: A Fully Dockerized Adversarial Robustness Assessment Framework for LiDAR-Based 3D Perception Models

The authors present a Dockerized framework connecting LiDAR models through HTTP interfaces. It supports adversarial testing of segmentation and object detection, enabling systematic vulnerability assessment across different model architectures.

Original source View record →

Research papersarXivPosted 2026.10.29EnglishPreprint

Assuring Learning-Enabled Cyber-Physical Systems with Foundation Models: A Neuro-Symbolic Agenda for Testing, Verification, and Runtime Assurance

This research agenda connects neuro-symbolic architectures, specification-based testing, compositional verification and runtime assurance for learning-enabled cyber-physical systems, including autonomous vehicles. It emphasizes system-level evidence rather than isolated model evaluation.

Safety architecture Automated driving Runtime monitoring Safety assurance Scenarios Uncertainty & OOD Verification & validation

Original source View record →

Research papersarXivPosted 2026.10.26EnglishPreprint

Uncertainty Evaluation to Support Safety of the Intended Functionality Analysis for Determining the Applicability of Performance Targets Across Triggering Conditions in ML-Based LiDAR Object Detection

The study tests whether uncertainty thresholds selected in one condition preserve detection performance targets in another. Its LiDAR experiments emphasize frame dependence, independent sequence counts and retained correct detections when interpreting SOTIF evidence.

Data quality Perception SOTIF Uncertainty & OOD

Original source View record →

Technical articles세온이앤에스 (Seon E&S)2026.10.25한국어

How Automotive SPICE MLE differs from conventional software development: Machine-learning engineering through automated driving

Uses a pedestrian detector to explain ML requirements, architecture, training, testing and data management. Covers training/test-data roles, provenance, labels, bias and integration with conventional SWE processes through traceable work products.

Data quality Lifecycle ML assurance Verification & validation

Original source View record →

Research papersarXivPosted 2026.10.24EnglishPublication / acceptance: author-reported

C-DFF: Concept-Level Diagnostics for Mechanism Parity in Simulation-Based Validation of Autonomous Driving Perception

C-DFF compares the semantic concepts influencing a frozen system on paired real and synthetic images. The diagnostic identifies shifts in decisive visual evidence, complementing output-based metrics in autonomous-driving simulation validation.

Automated driving Perception Verification & validation

Original source View record →

Research papersarXivPosted 2026.10.20EnglishPreprint

Provisioning Robust and Efficient LiDAR Perception for Autonomous Driving

The authors combine corruption-sensitive training with network and point-sampling reductions to provision LiDAR detectors under latency and memory constraints. Experiments assess detection accuracy, corruption robustness and resource use together.

Safety architecture Automated driving Data quality Perception

Original source View record →

Research papersarXivPosted 2026.10.16EnglishPublication / acceptance: author-reported

Expand-and-Prune: Test-Time Candidate-Space Adaptation for End-to-End Autonomous Driving

Expand-and-Prune adapts trajectory candidates at test time without changing learned parameters. It adds intent-guided and temporally propagated candidates, then prunes using prediction uncertainty, with evaluations across multiple autonomous-driving planners.

Automated driving Scenarios Uncertainty & OOD

Original source View record →

Technical articlesTroy Technical2026.10.15日本語

Automotive AI explainer: ODD, architecture and evaluation evidence

Troy Technical’s Japanese synthesis distinguishes SAE automation levels from ODD and compares modular/end-to-end architectures, log replay, simulation and road evaluation. It connects standards including ISO/PAS 8800 with public research to explain differences between evaluation metrics and safety evidence.

Safety architecture ISO/PAS 8800 ODD Scenarios Verification & validation

Original source View record →

Research papersarXivPosted 2026.10.15EnglishPreprint

Can Autonomous Vehicles Keep Their Cool? A Survey of Thermal Side-Channel and Perception Attacks

This survey examines thermal-infrared perception spoofing and information leakage through compute thermal telemetry. It compares attack conditions and defenses, identifying residual risks and gaps in automotive evaluation and sensor-compute co-design.

Safety architecture Automated driving Perception SOTIF

Original source View record →

Technical articles세온이앤에스 (Seon E&S)2026.10.11한국어

When a safe state becomes an attack path: Conflicts between functional-safety error handling and security responses

Analyzes how safety reactions can conflict with security threats through CAN fault-confinement and ECU error-handling examples. Connects fault/attack evidence, timing budgets, safe-state transitions and joint design review.

Safety architecture Functional safety Runtime monitoring

Original source View record →

Research papersarXivPosted 2026.10.11EnglishPreprint

Verifying Autoware Controller Node by Combining Model Verification and Runtime Monitoring

The authors model Autoware trajectory control as a hybrid system and prove safety and liveness properties. Runtime monitoring checks compatibility between the verified abstraction and implementation behavior during numerical simulation.

Automated driving Runtime monitoring Verification & validation

Original source View record →

Research papersACM2026.10.07English

ProbGuard: Proactive Runtime Monitoring for LLM Agent Safety via Probabilistic Prediction

ProbGuard learns symbolic Markov models from agent traces to predict future safety violations and trigger interventions. Autonomous-driving and household-agent experiments examine advance warnings and the trade-off between risk reduction and task completion.

Automated driving Runtime monitoring Scenarios

Original source View record →

Per page

Automotive AI safety library

LiDAR-SHIELD: A Fully Dockerized Adversarial Robustness Assessment Framework for LiDAR-Based 3D Perception Models

Assuring Learning-Enabled Cyber-Physical Systems with Foundation Models: A Neuro-Symbolic Agenda for Testing, Verification, and Runtime Assurance

Uncertainty Evaluation to Support Safety of the Intended Functionality Analysis for Determining the Applicability of Performance Targets Across Triggering Conditions in ML-Based LiDAR Object Detection

How Automotive SPICE MLE differs from conventional software development: Machine-learning engineering through automated driving

C-DFF: Concept-Level Diagnostics for Mechanism Parity in Simulation-Based Validation of Autonomous Driving Perception

Provisioning Robust and Efficient LiDAR Perception for Autonomous Driving

Expand-and-Prune: Test-Time Candidate-Space Adaptation for End-to-End Autonomous Driving

Automotive AI explainer: ODD, architecture and evaluation evidence

Can Autonomous Vehicles Keep Their Cool? A Survey of Thermal Side-Channel and Perception Attacks

When a safe state becomes an attack path: Conflicts between functional-safety error handling and security responses

Verifying Autoware Controller Node by Combining Model Verification and Runtime Monitoring

ProbGuard: Proactive Runtime Monitoring for LLM Agent Safety via Probabilistic Prediction

Participation enquiry

Your enquiry has been sent.