PhD Position on Vision-Language-Action Model Evaluation for Autonomous Driving

This fully funded, four-year PhD position at TU Delft focuses on evaluating Vision-Language-Action (VLA) models for autonomous driving using closed-loop simulation methods that generate photorealistic sensor feedback after each action. The project addresses the limitations of existing open-loop benchmarks by capturing closed-loop interactions between a vehicle and its environment.

Achieving this requires novel view synthesis techniques to enable realistic simulation from previously unseen viewpoints, as well as integrating language as a modality for evaluation, reasoning, and interpretability. One goal of the project is to develop public datasets and benchmarks specifically designed for closed-loop VLA evaluation.

The research is embedded in an international academic–industrial collaboration and targets fundamental questions in end-to-end autonomous driving and neural view synthesis. Your work is expected to lead to publications at top-tier venues such as CVPR, ICCV, ECCV, NeurIPS or ICRA. You will have access to extensive compute resources at TU Delft, ranging from local GPU servers to large-scale HPC infrastructure. The PhD will be supervised by Dr. Holger Caesar.

Job requirements

• Interest in performing cutting edge research in an active and exciting research area.
• Strong academic record with a solid background in Machine Learning.
• Knowledge of Vision-Language-Action models and Novel View Synthesis techniques is a strong plus.
• Good programming skills (Python) and knowledge of deep-learning frameworks (PyTorch) are expected.
• Affinity towards turning complex concepts into real-world practice is desired.
• Ability to act independently, collaborate effectively with members of a larger team, and supervise Master students.
• Good English skills are required.

Faculty Mechanical Engineering

ME is a dynamic and innovative faculty with high-tech lab facilities and international reach. It’s a large faculty but also versatile, so we can often make unique connections by combining different disciplines.