Main Terms: AI Safety for Autonomous Vehicles

ABS (Anti-lock Braking System) – Prevents wheel lock-up during braking, ensuring stability and control.

AI Bias – Systematic errors in AI caused by biased training data, impacting safe decision-making.

AI Safety – Ensuring AI systems operate reliably, predictably, and without unintended harm.

ASIL (Automotive Safety Integrity Level) – ISO 26262 risk classification determining safety rigor (A–D).

Autonomy Levels (SAE J3016) – Classification of driving automation from Level 0 (manual) to Level 5 (full automation).

Actuator – A device that executes control commands (e.g., steering, braking), requiring safe design.

Algorithm Validation – Process of confirming that AI/ML models perform as intended under all relevant conditions.

Black Box Testing – Evaluating system behavior based only on inputs/outputs.

Brake-by-Wire – Electronic braking system replacing mechanical/hydraulic linkages.

Boundary Conditions – Edge-case operating scenarios testing safety margins.

Behavioral Safety – Ensuring an autonomous system behaves acceptably in all conditions.

Benchmarking – Comparing AI or safety system performance against established standards or peers.

Computer Vision (CV) – AI that interprets visual data for perception tasks.

Cybersecurity – Protection from malicious attacks that could compromise safety.

COTS (Commercial Off-The-Shelf) – Pre-made components requiring validation in safety systems.

Certification – Formal approval verifying compliance with safety standards.

Collision Avoidance System – Technology that predicts and prevents crashes.

Control System – Hardware/software managing vehicle dynamics, safety-critical in autonomy.

Data Annotation – Labeling training data for AI/ML perception systems.

Data Drift – Change in input data over time, degrading AI performance and safety.

Deep Learning – Neural network-based ML technique central to perception but harder to validate.

Driver Monitoring System (DMS) – Monitors driver alertness and readiness to take control.

Duty Cycle – Duration and conditions under which off-highway machinery operates, relevant for reliability.

Edge AI – Running AI models directly on vehicles/machines with limited computing resources.

EHSR (Essential Health and Safety Requirements) – Safety requirements set by EU regulations like the Machinery Regulation.

Electronic Control Unit (ECU) – Embedded system managing functions such as braking, steering, or sensors.

End-to-End Learning – Training AI from raw input to output actions; difficult to validate in safety contexts.

Error Propagation – Chain reaction of small errors leading to unsafe outcomes.

Fail-Safe – System design ensuring safe operation when a fault occurs.

Fault Injection – Testing safety by deliberately introducing faults.

FMEA (Failure Modes and Effects Analysis) – Method for identifying potential failure points.

FMEDA (Failure Modes, Effects and Diagnostic Analysis) – Extension of FMEA adding diagnostics and fault detection.

Functional Safety (FuSa) – Ensuring systems remain safe under hardware/software faults (e.g., ISO 26262, IEC 61508).

Gap Analysis – Identifying gaps between current practices and required safety standards.

Generalization – AI’s ability to handle unseen scenarios safely.

Guardrails (AI) – Constraints to prevent unsafe or unintended AI behavior.

Ground Truth – Verified reference data used for validating AI predictions.

HARA (Hazard Analysis and Risk Assessment) – ISO 26262 process to identify hazards and assign ASIL levels.

Hazardous Event – Situation where a system fault or limitation could cause harm.

Human-in-the-Loop (HITL) – Approach where human supervision or override is part of safety.

HMI (Human-Machine Interface) – Interaction between operator and system, crucial for safe usability.

Hardware-in-the-Loop (HIL) – Simulation testing method connecting real hardware with virtual environments.

IEC 61508 – Generic functional safety standard for electrical/electronic/programmable systems.

IEC 62061 – Functional safety of machinery using electrical and electronic control systems.

ISO 13849 – Safety of machinery control systems, defines Performance Levels (PL).

ISO 19014 – Functional safety for earth-moving machinery, adapted from ISO 26262.

ISO 21448 (SOTIF) – Safety of the Intended Functionality, focused on AI and perception.

ISO 25119 – Functional safety for agricultural/forestry machinery (AgPL).

ISO 26262 – Automotive functional safety standard for E/E systems.

Jerk – Sudden change in acceleration, monitored for passenger comfort and safety.

J3016 (SAE) – Standard defining levels of driving automation.

Justification Argument – Structured reasoning used in safety cases to prove compliance.

Knowledge Representation – How AI encodes world knowledge for decision-making.

Key Safety Indicators (KSI) – Metrics for monitoring system safety performance.

Kalman Filter – Algorithm used for sensor fusion in autonomous systems.

LIDAR (Light Detection and Ranging) – Sensor technology providing 3D perception.

Latency – Time delay between input and system response, critical in safety.

Learning-based Control – AI-driven adaptation of control strategies, requiring strict safety validation.

Lifecycle (Safety) – End-to-end process from concept through decommissioning defined by safety standards.

Machine Learning (ML) – AI methods enabling systems to learn from data.

Machinery Regulation (EU 2023/1230) – EU regulation introducing AI-focused EHSRs.

Misuse Case – Analysis of how improper use may create hazards.

Mission Profile – Expected operating environment and usage for safety validation.

Monitoring Function – Safety mechanism that supervises normal functions to detect faults.

Neural Network – AI model inspired by the brain, central to modern perception systems.

Non-determinism – When a system’s behavior cannot be predicted exactly, raising safety challenges.

Nominal Operation – Standard, fault-free system functioning.

Normalization – Data processing step ensuring consistent AI inputs, reducing risk of errors.

Obstacle Detection – AI function identifying hazards in the environment.

OEDR (Object and Event Detection and Response) – Key capability defined in SAE J3016 for autonomy.

Operational Design Domain (ODD) – Conditions where an autonomous system is intended to operate.

Overfitting – AI model too tailored to training data, failing in new scenarios.

Performance Level (PL) – Safety classification in ISO 13849 for machinery control systems.

Perception System – AI + sensors enabling a machine to understand its environment.

Predictive Maintenance – Using AI to anticipate failures and prevent breakdowns.

Proven-in-Use Argument – Safety case using field data to justify reliability.

Probabilistic Safety – Approach considering probabilities of failures and unsafe outcomes.

Quality Assurance (QA) – Systematic processes ensuring safety and reliability in engineering.

Quantitative Risk Assessment – Assigning numerical values to safety risks.

Quasi-static Test – Testing under slow, controlled conditions to evaluate system behavior.

Redundancy – Multiple independent elements to ensure safety if one fails.

Risk Assessment – Systematic process for identifying and evaluating hazards.

Rollback Strategy – Safety mechanism to revert to a known safe state.

Runtime Monitoring – Real-time checks ensuring safety during operation.

Safety Case – Structured argument demonstrating that a system is acceptably safe.

Safety Integrity Level (SIL) – Risk reduction levels in IEC 61508 and IEC 62061.

Safety Mechanism – Function designed to detect and control faults or hazards.

Scenario-based Testing – Validating AI/autonomy in diverse real and simulated environments.

Sensor Fusion – Combining data from multiple sensors for more reliable perception.

SOTIF (Safety of the Intended Functionality) – Ensuring safe operation under performance limitations, not just faults.

Simulation Testing – Virtual environments for validating AI safety.

Testing Coverage – Measure of how thoroughly system functions are tested.

Training Data – Dataset used to teach AI models, critical for safety.

Transition Demand – Request for a driver/operator to take over from autonomy.

Traceability – Ability to link safety requirements to implementation and testing.

Telemetry – Remote monitoring of vehicle/machine state for safety and maintenance.

Uncertainty Quantification – Measuring confidence in AI outputs to improve safety.

Use Case – Specific operational scenario a system must handle safely.

Update Over-the-Air (OTA) – Wireless software updates requiring safe validation.

Validation – Ensuring a system meets its intended safety requirements.

Verification – Confirming a system has been built according to its specifications.

Virtual Validation – Use of simulation to assess AI/autonomous safety.

V&V (Verification & Validation) – Combined process to ensure systems are correct and safe.

Watchdog Timer – Safety mechanism resetting systems if they stop responding.

Worst-Case Scenario Analysis – Testing system robustness in extreme conditions.

Wireless Connectivity – Communication channel introducing both opportunities and cybersecurity risks.

Explainability (XAI) – Making AI decision-making transparent and interpretable for safety.

XML Safety Data Exchange – Structured format for exchanging safety-related configuration data.

Yielding Behavior – Autonomous system’s ability to give right-of-way safely.

Yellow Warning State – Intermediate warning state before critical failure.

YOLO (You Only Look Once) – Real-time object detection algorithm used in perception.