Electromyography (EMG) is a common method to estimate users’ intended motion in exoskeleton robots. Yet, it is highly susceptible to noise, and the process is complex due to its dynamic nature. While deep neural networks (DNNs) have been utilized to address these challenges, their high computational demands have necessitated costly GPU hardware and posed barriers to portable medical applications. This study introduces a novel approach employing a recurrent structure based on radial basis function neural networks (RBFNN) alongside a type-2 fuzzy inference system to mitigate computational costs while handling increased complexity. Integrating RBFNN and fuzzy inference enhances the model’s ability to process data with fewer parameters and reduced mathematical computations, enabling real-time practical solutions. An impedance controller manages robot joints, while a fuzzy system determines the control parameters. Most fuzzy model parameters are selected via a data-driven approach facilitated by a reinforcement learning (RL) algorithm acting as an expert and augmented by adaptive rules within the control cycle for the remaining parameters. Patients with mobility disorders can use video games to enhance the effectiveness of their rehabilitation. Also, video games can aid patients by customizing the controller’s parameters to suit their needs. A hybrid learning framework that combines supervised learning and RL has been developed and tested to determine fuzzy model parameters through a car driving task. The framework achieves precision and generalization similar to convolutional neural networks with similar normalized root mean square error (NRMSE) and correlation coefficient but fewer computational costs and floating-point operations per second.

Keywords