Deep learning is central to achieving SAE Level 4 and 5 autonomy, powering perception, decision-making, and control in self-driving vehicles. This keynote explores the intersection of data management, neural networks, and real-time perception in building scalable, reliable autonomous systems. Topics include the role of high-quality, diverse data, the design of robust deep learning models, and the integration of multimodal sensor inputs. We’ll highlight key challenges such as edge deployment, interpretability, and simulation-to-reality transfer. Attendees will gain insights into the deep learning advancements enabling safe, adaptable, and fully autonomous driving in complex real-world environments.

During this keynote, you will:

• Understand how deep learning architectures and sensor fusion enable robust real-time perception in dynamic driving environments
• Learn strategies for managing and leveraging large-scale, diverse datasets to train safe and adaptable autonomous systems
• Explore cutting-edge techniques like self-supervised learning and simulation-to-reality transfer for scalable AV development

• What methodologies best ensure dataset completeness and class balance across long-tail driving scenarios in supervised and self-supervised pipelines?
• How can spatiotemporal perception models fuse asynchronous, multimodal sensor data to achieve low-latency, high-accuracy environmental understanding?
• What are the current best practices for compressing and optimizing deep neural networks for real-time inference on automotive-grade edge hardware?
• How can simulation environments be used to close the domain gap between synthetic and real-world data for perception and planning models?
• Which frameworks or methods enable interpretability and functional safety validation of black-box deep learning components in AV stacks?

Presenting the results of our World Cafés: The moderators summarize the key takeaways of their World Café Round Tables on the main stage and share the main findings with all attendees.