Upper Limb Modeling for Delay-Compensated Motion Prediction using Dilated Convolution Based on sEMG and IMU Sensor Fusion

Accurate and continuous prediction of human motion intention is essential for practical upper limb exoskeletons, particularly to compensate for delays caused by various factors. However, most prior studies focus on predicting individual joint angles and mapping them to robot joint motion. This approach limits generalization across human motions and restricts applicability to robots with different kinematic structures. To address these limitations, we propose a novel upper limb modeling method that predicts wrist positions from the predicted orientations of the upper arm and forearm. This allows our method to handle arbitrary arm movements and to be applied to robots with different kinematic structures. The modeling involves motion prediction using a deep neural network with a sensor fusion structure that combines surface electromyography (sEMG) and inertial measurement unit (IMU) signals. A significant feature of the prediction model is the incorporation of dilated convolution to extract spatial correlations between sensors while keeping the model complexity low. In the offline experiments, the model maintained its prediction performance under various motion speeds, untrained motions, and rotational changes in sensor placement. These experiments demonstrate the model’s generalizability and robustness. In real-time experiments with a motor delay of 100 ms, the method reduced tracking error by up to 78.71% over a one minute of arbitrary movement compared with no motion prediction. The experiments also demonstrate that our method is applicable to robots with different kinematic structures and that it enables the robot’s end-effector to accurately track the human wrist by compensating for the motor delay.