Safety-Guided Deep Reinforcement Learning for Magnetic Microrobot Navigation and Precision Drug Delivery in Complex Vascular Networks

Magnetic microrobots are a promising approach for targeted drug delivery because they can be remotely guided toward localized therapeutic regions while reducing non-specific drug distribution. However, autonomous navigation in vascular environments remains difficult because of branching vessel structures, wall collision risks, misleading blood flow directions, target-location uncertainty, actuation disturbance, and restrictive vessel geometry. This study proposed a Safety-Guided Proximal Policy Optimization Deep Reinforcement Learning framework for magnetic microrobot navigation and target-zone drug delivery in simulated vascular networks. The proposed method combines a PPO-based navigation policy with a local magnetic safety-guidance layer that evaluates candidate actions, rejects unsafe movements, and selects vessel-aware, target-directed steering actions. A two-dimensional branching vascular environment was developed with randomized target placement, stochastic blood flow disturbance, collision constraints, misleading non-target branches, and automatic drug release upon entry into the target delivery radius. The framework was evaluated under progressively challenging scenarios, including standard conditions, moderate disturbance, high disturbance, and severe narrow vessel stress. The proposed method achieved 97.50% success under moderate disturbance, 87.50% success under high disturbance, and 69.17% success under severe narrow-vessel stress, showing realistic performance degradation as the environmental difficulty increased. Compared with the centerline, greedy direct, passive flow-following, and random baselines, the proposed framework maintained a stronger target-delivery performance and lower collision rates. These findings demonstrate that safety-guided reinforcement learning can improve the robustness of magnetic microrobot navigation in constrained vascular drug delivery simulations.