Abstract
TransSeptal Puncture (TSP) is a key step in many minimally invasive cardiac procedures, enabling access to the left atrium by crossing the interatrial septum from the right atrium. The task demands extreme precision, as excessive force may cause cardiac tamponade. Robotic platforms can improve precision and repeatability, but most of the existing systems are designed for training or teleoperation rather than autonomous execution, and do not incorporate adaptive impedance modulation. This work presents a learning-from-demonstration framework based on probabilistic impedance modeling to investigate the feasibility of learning and embedding expert force-regulation strategies for autonomous execution of the TSP under clinician-defined targets. Specifically, Gaussian Mixture Models, trained on teleoperated demonstrations, capture the relationship between contact force and operator stiffness, enabling real-time modulation of impedance through Gaussian Mixture Regression. The adaptive controller was deployed on a 7-DoF robotic platform and validated on fossa ovalis phantoms of varying compliance and on ex vivo porcine tissue. Performance was evaluated using clinically relevant metrics: Needle Puncture Force (NPF), Tenting Distance (TD), and Needle Stopping Space (NSS). Compared to fixed-stiffness baselines, the proposed controller reduced NPF by up to 35%, maintained TD within safe limits, and limited NSS below 1 mm in nominal anatomies, with consistent performance also observed under extreme phantom anatomies and ex vivo tissue. These findings demonstrate that embedding human-like impedance modulation can enable safe, anatomy-aware autonomous control, advancing robotic TSP toward clinical feasibility.

This publication has 38 references indexed in Scilit: