Autonomous Soft Robotic Guidewire Navigation via Imitation Learning

TL;DR

This work tackles autonomous navigation of soft robotic guidewires in intracranial aneurysm treatment where fluoroscopic feedback provides only partial visual information. It presents a transformer-based end-to-end imitation-learning framework with goal conditioning, relative motor outputs, and automatic contrast-dye management, formalized as $\pi_{\theta}(a_{t+1:t+K}^{motor}-a_t^{motor}, a_t^{contrast}|o_t)$. The policy leverages a CNN–Transformer architecture with a goal-distance feature map and outputs a motor-action chunk of size $K$ along with a contrast-injection decision, trained with $l_{L1}$ and $l_{BCE}$ losses (total $L=l_{L1}+0.5\,l_{BCE}$); it comprises 95.06M parameters and runs at 25 fps. Trained on 647 demonstrations across 36 modular mazes, the method generalizes to unseen geometries, achieving 83% success on novel layouts and outperforming diffusion, MLP, and centerline baselines, with ablations validating the necessity of recovery data, goal representation, action representation, and contrast strategy. These results suggest a viable path toward bench-top training translating to real-world endovascular interventions with limited additional demonstrations.

Abstract

In endovascular surgery, endovascular interventionists push a thin tube called a catheter, guided by a thin wire to a treatment site inside the patient's blood vessels to treat various conditions such as blood clots, aneurysms, and malformations. Guidewires with robotic tips can enhance maneuverability, but they present challenges in modeling and control. Automation of soft robotic guidewire navigation has the potential to overcome these challenges, increasing the precision and safety of endovascular navigation. In other surgical domains, end-to-end imitation learning has shown promising results. Thus, we develop a transformer-based imitation learning framework with goal conditioning, relative action outputs, and automatic contrast dye injections to enable generalizable soft robotic guidewire navigation in an aneurysm targeting task. We train the model on 36 different modular bifurcated geometries, generating 647 total demonstrations under simulated fluoroscopy, and evaluate it on three previously unseen vascular geometries. The model can autonomously drive the tip of the robot to the aneurysm location with a success rate of 83% on the unseen geometries, outperforming several baselines. In addition, we present ablation and baseline studies to evaluate the effectiveness of each design and data collection choice. Project website: https://softrobotnavigation.github.io/

Figures (8)

• Figure 1: (Top) Commercial guidewire and microcatheter for neurovascular intervention next to a small-scale soft robotic guidewire (under development) and our tool. (Bottom) Illustration of a soft robotic guidewire inside the vessels in the Circle of Willis (neurovascular structure). Here, we deploy a 3d-printed soft robotic guidewire in a 2D projection of various vascular geometries.
• Figure 2: To control the soft robotic guidewire, a user inputs force commands through a teleoperated control handle. These forces are proportionally mapped to the bending and translation velocities of the robot, achieved by the syringe pump and translation drive, respectively.
• Figure 3: First, we reserve a certain set of bifurcations and branches for the training set. From these sets of blocks, we choose a subset of combinations for the training set and a different subset for the rearranged test set. The novel test set is formed by a new set of bifurcations and branches.
• Figure 4: Proposed architecture for autonomous navigation. The static goal representation and live fluoroscopic image are passed to the autonomous policy. The policy outputs a sequence of relative actions and a binary variable indicating whether to inject contrast or not.
• Figure 5: Distance to the goal aneurysm boundary at the end of the trial for each ablative model. Each model was evaluated three times per goal aneurysm in each of the three rearranged and three novel geometries. Each individual bar represents 18 trials.