Autonomous driving relies on robust AI systems that require extensive training with vast amounts of data. However, data volume alone is not enough—its quality is equally critical. High-quality data must cover a balanced and relevant set, accurately representing the entire ODD of the vehicle. To meet these requirements, various data acquisition methods can be employed: recording real-world driving scenarios, augmenting existing data, and generating synthetic data through simulations or generative AI models. This talk will explore best practices and the potential of emerging technologies for efficient data acquisition. When recording large volumes of real-world driving data, trigger functions on edge devices are essential to capture only relevant information, optimizing bandwidth and storage usage. Data augmentation techniques allow for the creation of variations of existing scenes and simulation of rare edge cases, while generative AI offers unprecedented possibilities, enabling the generation of vast datasets through world models trained on large-scale internet data. By addressing these challenges, we aim to outline a roadmap for the efficient development and deployment of data-driven technologies in autonomous vehicles.

In this session, you will learn more about:

• Best practices for data-driven development
• Novel techniques for large-scale data needs such as augmentation and generative AI
• Approaches for combining multiple data sources, including real-world and generated data