Requirements on training data quantity for robust human arm posture prediction form pressure sensors

Abstract— Objective: This study investigates whether machine-learning models can be efficiently trained to predict human arm motion during horizontal drilling using pressure distribution data from a power-tool handle while maintaining high precision and robustness against varation of pressure distribution. Methods: Three participants perform horizontal wood drilling. Handle-integrated thin-film pressure sensors record pressure distribution. Arm joint motion (radial abduction, elbow flexion, shoulder flexion) is captured as ground truth using an IMU-based motion capture system (Xsens). For the arm posture prediction, Deep neural networks (DNN) and sparse Gaussian process regression (SGPR) models are trained. The influence of training data quantity is analyzed by varying the number of trials and datapoints per trial. Robustness is evaluated by the scatting of the prediction error across ten grouped train–test splits. Results: Accurate arm motion prediction was achieved for all joint angles. DNN models reached average RMSE values of 2.51° (radial abduction), 8.57° (elbow flexion), and 9.72° (shoulder flexion), with standard deviations below 1.6°. ... mehrRelative errors remained within 7–12% of the motion range. Increasing the number of independent trials significantly reduced RMSE, whereas datapoints per trial showed minimal influence beyond ~100 samples. SGPR models showed consistently lower precision and robustness than DNN models. Conclusion: The study demonstrates that sufficient quantity of training data enables robust and precise prediction of human arm posture during horizontal drilling from pressure sensor data. Maximizing the number of independent trials represents the key factor to ensure prediction robustness against varation of pressure distribution under constrained data volume.