Abstract
Accurate calibration-free recognition of upper-limb motion intention from surface EMG (sEMG) is essential for practical rehabilitation robotics. We propose a unified framework that couples self-supervised temporal-spectral pretraining with adversarial domain alignment. First, a masked time-frequency modeling objective learns subject-invariant features by reconstructing occluded spectrogram patches. Then, during fine-tuning, a gradient-reversal domain branch aligns latent distributions across subjects/datasets while a label head preserves class separability. Evaluations on NinaPro DB2 and CapgMyo DBa under leave-one-subject-out (LOSO) and cross-dataset protocols demonstrate consistent gains over traditional and deep baselines. In LOSO, the proposed method achieved 89.4% and 86.9% accuracy on DB2 and DBa, respectively; in cross-dataset transfer, it reached 82.1% (DB2 → DBa) and 83.5% (DBa → DB2). Ablations confirm that self-supervision and adversarial alignment are complementary, and robustness analyses show graceful degradation under noise and channel perturbations. These results indicate a scalable, portable, and plug-and-play sEMG interface suitable for clinical and home-based upper-limb rehabilitation.