Sample-Efficient Learning with Online Expert Correction for Autonomous Catheter Steering in Endovascular Bifurcation Navigation

A research team from multiple institutions has introduced a novel reinforcement learning (RL) framework designed to make autonomous surgical catheter steering significantly more efficient and robust. The paper, accepted by IEEE ICRA 2026, tackles the critical challenge of training AI for delicate endovascular procedures, where conventional RL suffers from sparse rewards and struggles to generalize. The proposed solution integrates three key components: a segmentation-based module for real-time state feedback, a fuzzy controller for intelligent orientation adjustment at vascular branches, and a structured reward generator that incorporates expert surgical knowledge to guide the AI's learning process more effectively.

By leveraging this 'online expert correction,' the framework dramatically reduces the inefficiency of random exploration typically required in RL. Experimental validation on a robotic platform using a transparent vascular phantom demonstrated concrete results: the system achieved policy convergence in just 123 training episodes, a 25.9% reduction compared to the baseline Soft Actor-Critic (SAC) algorithm. Furthermore, it reduced the average positional error to 83.8% of the baseline's performance. This advancement is particularly impactful for navigating anatomically challenging bifurcations, a common and difficult scenario in endovascular surgeries like aneurysm repairs or stroke interventions, paving the way for more reliable and accessible robot-assisted remote surgery.