Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation

Ischemic stroke caused by large vessel occlusion requires rapid mechanical thrombectomy, a procedure where surgical devices are navigated through tortuous human vasculature to remove cerebral blood clots. Timely intervention is critical because patient outcomes degrade rapidly with every minute of delay. However, only a small fraction of eligible patients receive mechanical thrombectomy due to geographic barriers, transfer delays, and a shortage of specialized neurointerventionists. Autonomous robotic systems powered by artificial intelligence can potentially bridge this gap by enabling less experienced clinicians in local hospitals to perform navigation safely. While reinforcement learning models have shown promise in simulation, previous controllers typically struggle with long navigation tasks across complex, unseen patient vasculatures.

This study investigates Progressive Experience Fusion (PEF), a training framework designed to construct a single multi-task agent capable of handling complex endovascular procedures across diverse vascular geometries. Instead of training isolated models for individual anatomy, PEF first groups experience replay data across different vascular shapes within specific subtasks, and then fuses these experiences across multiple tasks to train a unified TD-MPC2 world model. This world model learns a compact internal representation of the environment dynamics, allowing the controller to evaluate candidate action sequences virtually.

To improve navigation over long and difficult paths, the authors introduce an adaptive Model Predictive Path Integral planning horizon. This heuristic adjusts the planning horizon based on residual action-sequence dispersion, allocating more computational planning depth when the agent encounters uncertainty. Furthermore, the framework incorporates a clinically motivated fine-tuning stage, where the general agent interacts briefly with patient-specific simulation models prior to the actual procedure to adapt to completely unseen anatomies.