One of the hardest problems in medical robotics is bridging the gap between simulation and the operating room. In our work on autonomous endovascular navigation, we learned that a high-fidelity simulator is not enough — you need a systematic pipeline for sim-to-real transfer.

The simulation trap

It is easy to overfit to simulation dynamics. The physics engine may model fluid dynamics with impressive accuracy, but tissue compliance, patient variability, and imaging artefacts are fundamentally different in vivo. We addressed this through three strategies:

Domain randomisation — randomising tissue stiffness, imaging noise, and catheter properties during training
Curriculum learning — starting with simple anatomies and progressively introducing complexity
Real-world fine-tuning — a short phase of supervised learning on real biplane X-ray sequences

Key results

Our approach reduced the sim-to-real gap by 40% compared to naive transfer. The full methodology is detailed in our CathSim paper, but the practical lesson is simpler: invest as much in the transfer pipeline as you do in the simulator itself.

“The simulator is a hypothesis, not a guarantee.”

Next steps

We are now extending this work to multi-agent scenarios where multiple guidewires must coordinate in a shared vascular tree. The combinatorial complexity is significantly higher, but the principles remain the same.