Smoothness-Aware Hierarchical Planning and RCM-Constrained Execution for Robot-Assisted Neuroendoscopic Access: A Patient-Specific Computational Study

Robot-assisted neuroendoscopic access requires obstacle-responsive planning while maintaining a fixed cranial entry-point remote center of motion (RCM). We computationally evaluated a Hierarchical Planning-and-Execution Framework combining a proximal policy optimization (PPO) local planner generating Cartesian velocity commands with analytical RCM-constrained execution. Patient-specific scenes were reconstructed from multimodal imaging; seven patient anatomies supported development and within-cohort validation, and three were held out from training. Smoothness-Aware PPO+RCM was obtained by fine-tuning the original policy with an action-transition penalty. Under navigation-map perturbations, Smoothness-Aware PPO+RCM achieved a target error of 0.616 ± 0.236 mm, comparable to PPO+RCM (p = 0.723), while mean angular change decreased from 21.66 ± 3.87° to 1.07 ± 1.56° (p < 0.001). Full-trajectory time-averaged RCM error was 4.68 ± 4.70 mm; post-alignment RCM performance remained comparable to PPO+RCM (p = 0.901). Composite Constrained Task Success was 33/50 (66%). Across three held-out patients, mean target error ranged from 0.58 to 0.78 mm, post-alignment RCM error from 1.88 to 2.34 mm, and 17/30 scenes met the composite criterion, with no direct geometric intersections with simplified risk models. These findings support improved computational trajectory regularity with preserved targeting, geometric clearance, and RCM performance. Physical robot validation remains necessary.