Five-Layer MLOps Architecture Enables Collective Learning for Self-Driving Fleets
A blueprint for connected vehicles to learn together, detect black swan events, and stay safe over time.
Automated driving systems (ADS) face a fundamental challenge: how to continually assure safety and performance in dynamic, open-world environments where rare or unforeseen scenarios emerge. Traditional static validation fails, but machine learning offers adaptation through operational data. The key advantage of fleets vs. human drivers is collective data sharing across vehicles—even across different operators. Lampe and Eckstein's five-layer MLOps architecture capitalizes on this, providing a structured approach for collaborative learning and continuous improvement.
The architecture defines five layers with clear responsibilities and interactions—from data ingestion and feature engineering to model training, deployment, and feedback loops. A unique contribution is the integration of multi-level self-assessments that help detect and mitigate edge cases, including catastrophic 'black swan' events. The paper (8 pages, 6 figures) serves as a practical blueprint for fleet operators, OEMs, and MLOps engineers to design systems that scale safety assurance through collective intelligence across connected vehicles.
- Proposes a five-layer MLOps architecture specifically for connected automated driving systems (ADS) to enable continuous learning across fleets.
- Addresses detection and reduction of edge cases and black swan events through integrated multi-level self-assessments.
- Provides a conceptual blueprint (8 pages, 6 figures) for fleet operators to implement collective MLOps processes across multiple vehicle fleets.
Why It Matters
Enables autonomous vehicle fleets to continuously improve safety and performance through shared learning and edge-case detection across operators.