The goal of this project is to develop a reinforcement learning agent for autonomous guidewire navigation in simulated vascular anatomies. The work will build on the open-source stEVE_training framework.
Unlike existing approaches that mainly use positional information, the agent will receive force-based observations generated by interactions between the guidewire and the vessel wall. The aim is to investigate whether force signals can be used to navigate safely toward predefined targets.
The student will adapt the simulation environment, define suitable force-based reward functions, train the agents, and evaluate its performance. Results will be compared with a position-based baseline using metrics such as navigation success, procedure time, applied forces, vessel-wall contacts, and generalization to unseen anatomies.
Tasks
- Extend
stEVE_training with force-based observations
- Adapt the reinforcement learning environment and reward function
- Train and optimize the navigation agent
- Evaluate safety, performance, and generalization
- Compare the results with a position-based approach
Requirements
- Master’s student in Medical Engineering
- Good Python programming skills
- Excellent knowledge of machine learning and reinforcement learning
- High interest in medical robotics and simulation
