Estimating IMU signals from surface EMG using physics-informed and domain-adaptive neural networks

We propose a physics-informed, domain-adaptive neural network architecture for estimating body kinematics from surface electromyography (EMG) signals across heterogeneous human task domains. The proposed approach addresses the biomechanical inverse mapping problem by predicting inertial measurement unit (IMU) outputs at the window level directly from segmented and normalized EMG inputs, without requiring domain-specific tuning or retraining. The model employs a shared convolutional feature extractor followed by two branches: a primary regression head that predicts mean IMU signals and a domain-classification head trained adversarially using a gradient reversal layer. This architecture encourages the learning of task-invariant representations while preserving discriminative capacity for accurate motion prediction. To enforce biomechanical plausibility, we introduce a physics-informed loss term that penalizes the first derivative of the predicted acceleration (i.e., jerk), thereby reducing unrealistic dynamics. The model was evaluated using EMG-IMU data collected across five distinct task categories with varied sensor configurations. The proposed approach achieved a root mean squared error of 0.42 ± 0.11 m/s² for acceleration and 6.38 ± 1.1°/s for gyroscope outputs across 57 IMU output channels, corresponding to approximately 23% and 2.7% of peak amplitudes, respectively. Domain prediction accuracy reached nearly 75%, with per-class precision, recall, and F1 scores varying as a function of domain complexity and the degree of overlap in EMG activation and motion patterns across task domains.