Abstract
The shoulder girdle is one of the most complex components of the upper limb, and when coupled with the arm, modeling and prediction become even more challenging. To address this, we propose a modular motion-planning framework that explicitly decouples the shoulder girdle from the arm. This separation simplifies modeling, enhances interpretability, and improves adaptability for rehabilitation scenarios. By independently targeting scapular dynamics, the framework enables better generalization of motion prediction across subjects. For human-motion modeling, a Transformer-based deep learning architecture is employed to capture nonlinear dependencies between joint angles and scapular motion. The model accepts joint-specific features as input and predicts shoulder girdle configurations, which are then integrated with arm trajectories to reconstruct complete upper limb motion. In the exoskeleton-mapping stage, a machine learning framework translates predicted human motion into the configuration space of a 6- Degree of freedom (DOF) rehabilitation exoskeleton. This ensures that generated trajectories are physically realizable and clinically suitable. By isolating human-motion prediction from robot mapping, the framework remains modular, scalable, and resilient to subject-specific variability, making it ideal for personalized rehabilitation. The methodology was evaluated by comparing Transformer-based predictions with both experimental data, reinforcement learning models and long short-term memory across multiple rehabilitation tasks. Quantitative analyses included statistical measures (F1-Score, ANOVA, T-test) and kinematic error metrics (RMSE, DTW). Results demonstrated that the Transformer model achieved higher accuracy and better temporal alignment with experimental trajectories. Combining shoulder-arm separation with Transformer-based learning provides an effective and clinically relevant solution for generating human-like motion in upper limb rehabilitation.