Are you a student eager to apply your theoretical knowledge and fresh perspectives to real-world challenges? At Saab, we believe that innovation thrives on new ideas, and your master thesis project could be the spark that ignites our next technological breakthrough.
We recognize the immense value that students bring to our company. Your academic rigor, combined with your enthusiasm for cutting-edge technology, allows you to approach problems with a unique and insightful lens. At Saab, you’ll have the opportunity to collaborate with experienced engineers and specialists, gaining invaluable practical experience while making a tangible contribution to our growth and development.
The purpose of this thesis project is to investigate how reinforcement learning (RL) can be used to accelerate classical motion planning methods for unmanned vehicles, and how large language models (LLMs) can support the development of such RL solutions through automated reward function design and refinement.
A common weakness of classical optimal motion planning algorithms is that they may require substantial computation time to converge to a feasible and optimal solution. Learning-based alternatives, which predict motion through forward passes of a neural network, can generate solutions quickly but generally provide no guarantees of convergence or optimality. Combining classical and learning-based approaches to planning is therefore a promising research direction, offering the potential to combine the reliability of optimization-based methods with the efficiency of learned models.
Reinforcement learning has demonstrated promising results in domains that can be accurately modeled in simulation. However, designing reward functions that lead to effective learning remains a challenging and often iterative process. Integrating an LLM-driven component into the reward design loop, capable of proposing and refining reward formulations based on training outcomes, may reduce the effort associated with reward engineering while improving policy performance and training efficiency. Such an approach could allow reward functions to be adapted automatically based on feedback from training runs, reducing the need for manual tuning.