Welcome to the National Transport Library Catalogue

Normal view MARC view

Towards zero bottlenecks for scaling autonomous driving

By: Series: Doctoral theses in mathematical sciences ; 2025:1Publication details: Lund : Lund University, 2025Description: 83 sISBN:
  • 9789181042993
Subject(s): Online resources: Notes: Härtill 5 uppsatser Dissertation note: Diss. (sammanfattning) Lund : Lunds universitet, 2025 Summary: This dissertation examines the main scaling challenges in autonomous driving development, discussing recent advances in the field while contributing specific solutions to key bottlenecks. The first challenge is the reliance on human labor, particularly for annotations. Here we make two key contributions: new techniques to extract additional value from existing annotations through future prediction (I), and an adaptation of vision-language learning to 3D automotive sensors that reduces dependence on explicit labels while maintaining interpretability (II). The second challenge concerns access to training data covering the full spectrum of driving scenarios. We address this data bottleneck through complementary approaches: releasing a diverse European driving dataset collected across multiple years and conditions (III), and developing a neural rendering method that enables scalable generation of realistic synthetic data (Iv). Finally, to enable scalable safety testing, we introduce a closed-loop neural simulator that transforms ordinary driving scenarios into challenging near-collision cases (v). Together with broader advances in the field, our contributions suggest a promising path toward scaling autonomous vehicle development.
Item type: Dissertation
No physical items for this record

Härtill 5 uppsatser

Diss. (sammanfattning) Lund : Lunds universitet, 2025

This dissertation examines the main scaling challenges in autonomous driving development, discussing recent advances in the field while contributing specific solutions to key bottlenecks. The first challenge is the reliance on human labor, particularly for annotations. Here we make two key contributions: new techniques to extract additional value from existing annotations through future prediction (I), and an adaptation of vision-language learning to 3D automotive sensors that reduces dependence on explicit labels while maintaining interpretability (II). The second challenge concerns access to training data covering the full spectrum of driving scenarios. We address this data bottleneck through complementary approaches: releasing a diverse European driving dataset collected across multiple years and conditions (III), and developing a neural rendering method that enables scalable generation of realistic synthetic data (Iv). Finally, to enable scalable safety testing, we introduce a closed-loop neural simulator that transforms ordinary driving scenarios into challenging near-collision cases (v). Together with broader advances in the field, our contributions suggest a promising path toward scaling autonomous vehicle development.