At Kognic, we believe that machines learn faster with human feedback. A few weeks ago, we celebrated the successful presentation of Tobias Hoek's master thesis, which exemplifies this principle. During his time with us, Tobias advanced the field of traffic scenario classification—a critical component in developing autonomous systems that are safe, reliable, and aligned with human expectations.
In the automotive industry, accurate scenario classification is essential for understanding complex driving situations and ensuring autonomous systems behave in ways that feel safe and consistent with human judgment. Tobias developed a supervised scenario classification approach that accurately categorizes various traffic scenarios, including ego lane identification—work that directly supports our mission to integrate cost-efficient, scalable human feedback into autonomy data pipelines.
Tobias' research utilized Graph Convolutional Networks (GCNs) for spatial aspects and Convolutional Neural Networks (CNNs) for temporal aspects. This powerful combination proved highly effective, achieving a serious error rate of less than 30%—results that demonstrate how the right combination of machine learning and expert human validation can push the frontier of annotation for autonomy forward.
His work sets a benchmark for future research in traffic scenario classification, opening new possibilities for how we validate and train autonomous systems. As annotation evolves from basic labeling to judging intent, ranking behaviors, and validating machine decisions against human expectations, research like Tobias' helps us stay at the forefront. We're proud of his contribution to Kognic and the broader autonomy community as we continue making our roads safer and more efficient.
Learn more about this research here