Self-learning of inverse kinematics for feedforward control of intracardiac robotic ablation catheters

This paper investigates the self-learning of inverse kinematics for the feed-forward control of a robot to position intracardiac catheters. Cardiac ablation is routinely performed to treat Atrial Fibrillation, and requires a catheter be accurately positioned in the heart, by hand or by a robot, under feedback control. This is typically a slow process and methods to reduce procedure times are needed. To investigate our proposed method, a robotic system to manipulate a standard intracardiac catheter was constructed. To safely develop our proposed learning system, a comprehensive dataset was collected using a magnetic tracking system to measure the catheter tip positions versus robot actuator positions. Initially, the robot began with no model of its kinematics. A Genetic Algorithm was used to decide on the next actuator sequence that would reduce the uncertainty in a Feedforward Neural Network (FFNN) based inverse kinematic model. An automated iterative process was followed where the robot would perform virtual experiments, to grow its knowledge of its inverse kinematics. After 791 learning cycles the final analysis revealed that the complete inverse kinematic relationship has been explored with the given constraints. A validation dataset indicated the learned FFNN model was able to predict x, y and z positions of the catheter tip to within ±0.17 mm, ±0.73 mm and ±0.62 mm, respectively. The robot successfully self-learned its inverse kinematic model using the proposed methodology. Future work is required to investigate the influence of disturbances on positioning accuracy.