Autonomous driving development faces persistent data challenges: identifying high-value “golden” data, optimizing the learning loop, and addressing the scarcity of diverse datasets for robust performance analysis. Golden data—only a small fraction of collected data—is truly valuable for training models effectively. Optimizing the learning loop requires analyzing model evaluation results to determine which data should be added next, making the cycle more efficient and targeted. Additionally, augmenting real-world data provides “apples to apples” data that enables consistent regression testing and broader coverage. Microsoft addresses these challenges through a dual approach: building in-house AI solutions and collaborating with partners to deliver innovative tools for mutual customers.

• Small Language Models (SLMs) enhance scene understanding by describing and classifying driving scenarios in natural language, improving the precision of golden data extraction.
• AI agents recommend training data based on model evaluation results, streamlining the learning loop.
• Cognata’s DriveMatrix platform augments real-world data enabling scalable and repeatable testing.