BCIwiki (bciwiki.com) — A robotic controller trained with reinforcement learning autonomously navigated cerebral vein phantoms under fluoroscopy to reach the delivery site for an endovascular brain-computer interface (BCI), succeeding in 14 of 20 physical runs, according to a preprint posted to arXiv on October 2, 2026 (arXiv:2610.03537). The authors, Harry Robertshaw, Weijie Qi, Nikola Fischer, Alejandro Granados, Thomas C. Booth and Sam E. John, describe it as the first demonstration of in vitro autonomous robotic navigation for endovascular BCI access in the cerebral venous system. The work has not been peer reviewed.
Endovascular BCIs avoid craniotomy, but the device has to be delivered precisely through cerebral veins whose anatomy varies widely. The team trained Soft Actor-Critic controllers in simulation for two sequential tasks running from the right internal jugular vein to the superior sagittal sinus, using geometric augmentation of a single training anatomy.
Navigation was tested on the training anatomy and on an anatomically unseen hold-out model, with 250 simulated episodes and 5 fluoroscopy-guided robotic runs per task-anatomy condition, or 1,000 simulated episodes and 20 physical runs in total. In simulation, Tasks A and B succeeded 85.6% and 98.4% of the time in the training anatomy, and 42.0% and 91.6% in the hold-out anatomy. Overall physical success was 70%, including 80% for Task B in the hold-out phantom.
The researchers also tested recurrent predictors designed to flag impending navigation failure online. In simulation, the predictors caught 99.3% to 100.0% of failures with false-alarm rates of 0.8% to 6.7%. In the physical runs, predicted risk rose before failed episodes, but some successful runs also showed elevated failure probabilities, indicating weaker calibration after transfer from simulation.
The authors say the results show autonomous cerebral venous access is feasible and that failure prediction could support human oversight. They identify generalization across anatomies and sim-to-real calibration as priorities before preclinical translation.