We are committed to shaping the future of automotive mobility by developing highly automated driving systems for both highway and urban areas. Our development teams in Germany and California work with state-of-the-art technologies to develop innovative and class-leading systems to provide our customers with the best experience possible. To master this challenge, we are looking for energetic and committed PhD students to conduct research within our Scene Reasoning and Prediction Team in Stuttgart.
The focus of your PhD thesis will be on the development of foundation models for motion prediction in autonomous driving. Especially in natural language processing, foundation models have achieved state-of-the-art performance on various benchmarks and have become essential tools for researchers, developers, and businesses looking to build natural language processing applications. Their strength is their ability to learn the structure and nuance of language, enabling them to be fine-tuned on specific tasks or domains. In contrast, current motion prediction algorithms are usually trained task-specific, e.g., for the task of trajectory prediction. How to leverage task-unspecific pre-training on large autonomous driving datasets, potentially also with different training tasks, and to then perform fine-tuning for motion prediction tasks is still an open research question.
The key challenge of your PhD thesis is therefore to develop a foundation model for motion prediction. This includes the development of a task-unspecific pre-training and the fine-tuning for different prediction tasks, e.g., trajectory prediction and intent prediction. For this, you are expected to survey the latest state-of-the-art in motion prediction and foundation models. You will make use of the latest machine learning techniques, such as Graph Neural Networks and Transformers.
Responsibilities:
Assessment of the current state-of-the-art in learning-based motion prediction and foundation models Development of a latent space representation and of training tasks to learn structure and nuance of traffic scenes from large-scale datasets during a task-unspecific pre-training Development of fine-tuning methods for different motion prediction tasks Implementation of corresponding training and evaluation concepts Integration of the developed methods into the AD software Evaluation of the developed methods in simulation, recorded driving data and real vehicles
The final thesis selection is made in close consultation with you, the university and us.