Scaling Zenseact's Stack Towards Autonomy on a Next-Generation DL Supercomputer

Explore a 25-minute conference talk from the 2025 GAIA Conference where Willem Verbeke, Product Owner at Zenseact, discusses the challenges and solutions in scaling autonomous driving technology. Learn how Zenseact is developing an autonomous driving stack based on deep learning models that perform end-to-end sensor and temporal fusion, requiring substantial computational resources that increase with in-vehicle computing capacity. Discover why manual annotation is insufficient for training these models and how pseudo-annotations and self-supervised learning offer solutions by leveraging large compute capacity to reduce human annotation efforts. Understand the emerging importance of Neural Radiance Fields (NeRFs) and 3D Gaussian splatting for simulating sensor data to evaluate model performance under adverse conditions. Verbeke, who transitioned from particle physics at CERN to deep learning engineering at Zenseact, shares insights from his journey leading the "Deep Learning Data Enrichment" area focused on automating annotation processes with large offline models and mining high-value data. Recorded at Svenska Mässan in Gothenburg, Sweden on April 11.