Predicting future gait movements in children can provide crucial input for controlling lower limb robotic exoskeletons. Unlike previous research that primarily focused on predicting gait patterns of typically developing individuals, this study targets pathological gait forecasting, which is more complex due to high intra- and inter-subject variability. In this study, a deep temporal model, named as PathoGaitNet, is developed by combining a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network to predict joint angle trajectories at the hip, knee, and ankle in pitch and roll directions. The open-access dataset consists of motion capture recordings of children aged 4–19 years with various neurological disorders, primarily cerebral palsy, collected using a VICON system at 120 Hz. Three different input window sizes (200 ms, 400 ms, and 600 ms, corresponding to 24, 48, and 72 timesteps) were tested while keeping the output window size fixed at 8.33 ms (single timestep). Results show that the CNN-LSTM consistently outperformed the standalone CNN and LSTM models in terms of Mean Absolute Error (MAE), Mean Squared Error (MSE), and Pearson Correlation Coefficient (PCC), achieving the best performance with a 72-timestep input window (MAE = 0.631, PCC = 0.996). Joint-wise analysis revealed that the model achieved higher prediction accuracy for knee and hip joints than the ankle, likely due to the lower variability and smoother dynamics in proximal joints. These findings suggest the feasibility of deploying deep learning-based single-step gait forecasting in real-time control systems for rehabilitation robotics.

This paper introduces a novel robust control strategy specifically designed for a one-degree-of-freedom (1-DoF) knee rehabilitation exoskeleton robot, focusing on position control. Our approach addresses several key challenges, including state constraints, parameter uncertainties, solid and viscous frictions, and external disturbances. We suggest a composite controller that combines a nonlinear controller with a linear state feedback controller in order to successfully address these issues. By utilizing a quadratic Lyapunov function, we derive the expression of the nonlinear control law and establish the Linear Matrix Inequality (LMI) conditions for computing the matrix gain of the linear controller, ensuring robust stabilization of the exoskeleton robot to the desired position. These conditions are derived through the application of advanced mathematical techniques such as the matrix inversion lemma, the Young inequality, the Schur complement, and the S-procedure lemma. Numerical results show the effectiveness of our suggested control law, demonstrating the continuous convergence of the intended angular position to the real one, even when external disturbances and uncertainties are present. Overall, our research not only presents a novel methodology for robust control but also delivers promising results that underscore the efficacy of our approach in achieving stable and reliable performance for 1-DoF knee exoskeleton robots.