Toward Safe Autonomous Robotic Endovascular Interventions using World Models

A research team from King's College London and University College London has published a groundbreaking preprint demonstrating the first validated autonomous navigation system for robotic mechanical thrombectomy (MT), a critical procedure for treating ischemic strokes. The system, detailed in the paper "Toward Safe Autonomous Robotic Endovascular Interventions using World Models," leverages a world-model-based framework built on the TD-MPC2 model-based reinforcement learning (RL) algorithm. This approach allows the AI to plan actions within a learned internal model of the environment, which proved crucial for navigating the highly variable and tortuous geometries of human cerebral vasculature. In simulation tests across diverse, held-out patient anatomies, the TD-MPC2 agent achieved a mean success rate of 58%, significantly outperforming the 36% rate of a standard Soft Actor-Critic (SAC) agent.

The team further validated the system's performance using physical, patient-specific vascular phantoms under real fluoroscopic guidance, bridging the gap between simulation and clinical application. In these in vitro experiments, the TD-MPC2 agent maintained a comparable 68% success rate while achieving superior navigation efficiency, measured by a lower path ratio, meaning it took a more direct route through the vessels. Critically, the AI-operated robotic catheter maintained a mean tip contact force of just 0.15 Newtons during navigation, which is an order of magnitude below the proposed 1.5N safety threshold for vessel wall rupture. This combination of improved success rates and demonstrably safe force profiles highlights the promise of world models for creating generalizable and trustworthy AI assistants for complex, life-saving endovascular surgeries.