Autonomous Robotic Navigation for Endovascular Brain-Computer Interface Access
This study addresses the challenge that cerebrovascular anatomical variations severely constrain the precise delivery of endovascular brain-computer interface devices. To overcome this, we present the first demonstration of autonomous robotic cerebrovascular navigation in vitro, proposing a path planning framework based on Soft Actor-Critic reinforcement learning. The approach integrates geometric data augmentation with simulation-to-real transfer techniques and incorporates an online failure prediction model to facilitate human supervision. In simulated tasks, the system achieves a success rate of 98.4%, while physical experiments yield an overall success rate of 70% with a failure detection rate exceeding 99%. This work validates the feasibility of autonomous endovascular navigation while highlighting current challenges in generalization capabilities.