Physically Constrained Dual-Branch Front-End Optimization for DDH-Oriented Surgical Robot Navigation

Surgical robot navigation in a restricted pelvic workspace requires accurate target localization, directional consistency and robot-feasible execution. This study proposes a physically constrained dual-branch front-end network (PCD-Net) for DDH-oriented navigation using public CT-derived pelvis geometries. PCD-Net maps a 21-dimensional input comprising the target geometry, current joint state, workspace bounds and constraint parameters to a base-frame target position, principal insertion direction and seven-joint correction vector for MoveIt2 and OMPL planning. Training combines supervised pretraining with direction consistency, joint limit, correction magnitude and workspace gap penalties. Evaluation comprised 15 complete simulation trials per method and target sequence-level fivefold cross-validation of 999 samples from five complete sequences. PCD-Net achieved a position error of 0.736±0.202 mm, a direction error of 0.428±0.130°, a planning time of 0.0162±0.0040 s and successful execution in all 15 trials. In cross-validation, the complete constraint setting produced the lowest joint correction MAE (0.755±0.034 rad) and temporal correction variation (3.270±0.589 rad) while maintaining sub-millimeter position and sub-degree direction errors. Removing the joint limit penalty increased the violation rate by 42.35%. These results support PCD-Net as a lightweight, planner-compatible front end that balances geometric accuracy and joint-level feasibility. All evidence is simulation-based; phantom experiments, physical robot validation and evaluation using clinically characterized DDH cases remain necessary before surgical translation.