PiMAN: A Physics-informed Motion Prediction Network using sEMG signal features for human movement parameters

Abstract

Early and accurate prediction of human movement parameters is critical for assistive robotics to synchronize effectively with a user’s intent. Surface electromyography (sEMG) signals offer a unique advantage by capturing neuromuscular activity prior to visible motion; however, existing model-based and model-free approaches often suffer from limited generalizability, delayed response, or poor biomechanical interpretability. To address these limitations, we propose PiMAN (Physics-informed Motion Anticipation Network), a deep learning framework that combines an attention-based bidirectional gated recurrent unit (BiGRU) architecture with physics constraints derived from the inverse dynamics. The model incorporates subject-specific anthropometric hyperparameters into the inverse dynamics formulation, enabling biomechanically consistent torque estimation across individuals. PiMAN predicts a comprehensive set of joint parameters, including angles, velocities, accelerations, external payloads, and torques, 48–96 ms before visible movement onset, from sEMG windows aligned with electromechanical delay range. This supports real-time control in assistive and neuroprosthetic systems. The model was trained and evaluated on five test subjects under three external load conditions (0 kg, 2 kg, and 4 kg), using both intra- and inter-subject scenarios. It achieved low RMSE ($\le$1.3) and high correlation (up to 0.93) across all outputs. Compared to purely data-driven baselines and physics-informed variants lacking attention, PiMAN consistently outperforms in joint torque and load estimation, particularly under higher-load conditions. In addition, PiMAN generalizes to temporally varying load transitions without retraining, and treats external mass as a continuous variable to facilitate seamless integration into inverse dynamics. These findings position PiMAN as a scalable, generalizable, and real-time-ready framework for anticipatory motion prediction in wearable assistive technologies.