Luca BertonAutomotive AI: Where Safety Requirements Are Non-Negotiable
Automotive AI development requires infrastructure at a scale and safety standard that few other industries demand. A single autonomous driving development program generates petabytes of sensor data, requires thousands of GPU hours for training, and must demonstrate safety across billions of simulated miles — all while meeting the strictest functional safety standards in any industry.
The Automotive AI Stack
Data Management
- Scale: A single test vehicle generates 1-20 TB/day of sensor data (cameras, LiDAR, radar, IMU, GPS)
- Fleet scale: 100+ vehicle fleet = 100-2,000 TB/day ingestion
- Data pipeline: Ingest → validate → annotate → curate → version → train
- Annotation: AI-assisted labelling with human verification for safety-critical ground truth
- Data versioning: Every training dataset must be reproducible for regulatory traceability
Training Infrastructure
- GPU clusters: 100-1,000+ GPU training clusters for perception models (object detection, semantic segmentation, 3D reconstruction)
- Multi-modal training: Fusing camera, LiDAR, and radar data requires specialised model architectures
- Continuous training: Models retrained as new edge cases are discovered from fleet data
- Experiment tracking: Thousands of training runs with different hyperparameters, architectures, and data subsets
Simulation
- Scenario generation: Billions of simulated driving scenarios to test AI behaviour in edge cases
- Sensor simulation: Physically accurate camera, LiDAR, and radar simulation for training and testing
- Hardware-in-the-loop (HIL): Test AI on actual automotive compute hardware in simulated environments
- Regression testing: Every model update must pass thousands of critical scenario tests before deployment
Regulatory Framework for Automotive AI
Key Regulations
- EU AI Act (Annex I): AI safety components in vehicles are high-risk, requiring conformity assessment
- UNECE WP.29: UN regulations for automated driving systems (ALKS — Automated Lane Keeping Systems)
- EU General Safety Regulation (2019/2144): Mandates ADAS features (AEB, lane keeping, driver monitoring) from July 2024
- ISO 26262: Functional safety standard for automotive — ASIL (Automotive Safety Integrity Level) classification for AI components
- ISO 21448 (SOTIF): Safety of the Intended Functionality — specifically addresses AI/ML limitations and edge cases
- Cyber Resilience Act: Applies to connected vehicle components with digital elements
📘 Book
Kubernetes Recipes
Practical guide for container orchestration and deployment — hands-on patterns you can use today.
View on Amazon →Infrastructure Requirements
- Petabyte-scale storage: Object storage (S3-compatible) with data lifecycle policies for the multi-year retention required by regulators
- High-throughput data pipelines: Ingest terabytes per day from global vehicle fleets with guaranteed delivery
- Reproducible training: Every model version must be reproducible from the exact training data and parameters used
- Safety validation pipeline: Automated safety testing integrated into CI/CD — no model deploys without passing safety regression tests
- Over-the-air (OTA) updates: Secure, verified model updates to vehicles in the field with rollback capability
automotive
ADAS
autonomous driving
AI infrastructure
type approval
UNECE
functional safety