Patient-Specific AI-Guided Robotic Intraoral Scanning Using Deep Learning Residual Trajectory Control for the Dentatron Dental Robot

Accurate trajectory tracking is essential for autonomous dental scanning in the presence of patient-specific anatomical variations and robotic nonidealities. This study proposes a patient-specific trajectory-tracking framework for the four-DOF Dentatron dental robot using a deep learning residual controller. Ten patient-specific maxillary geometries were extracted and parameterized from upper-jaw intraoral scans obtained from the publicly available Teeth3DS+ dataset. The selected cases represented variations in intercanine width, intermolar width, total arch length, palatal depth, arch form, crowding, spacing, and asymmetry. Sixteen scanning targets were defined for each patient-derived geometry and converted into smooth Cartesian trajectories sampled at 100 Hz, resulting in a dataset of 31,000 trajectory samples. A MATLAB/Simulink–Simscape model of the Dentatron was developed incorporating representative robotic and sensing nonidealities, including joint dynamics, mechanical backlash, an 80 ms command delay, disturbances, encoder quantization, measurement noise, and sensor bias. The residual controller estimates joint-position and joint-speed corrections using patient-specific geometric descriptors, trajectory variables, and robot-state information. Patient-wise partitioning was employed to evaluate generalization to complete maxillary geometries excluded from model training. On the unseen test set, the overall joint-position RMSE decreased from approximately 0.34° to 0.26°, corresponding to a 22.7% improvement, while joint-speed RMSE decreased from approximately 1.25°/s to 0.96°/s, representing a 23.0% improvement. The strongest compensation was observed for joints with the largest baseline tracking errors, particularly J2 and J3, whereas smaller improvements were obtained for joints with lower initial errors. All generated patient-specific trajectories remained outside the adopted singularity region. These results demonstrate the feasibility of residual deep learning for improving patient-specific Dentatron trajectory tracking under simulated robotic nonidealities while maintaining kinematic feasibility.