Recent studies have demonstrated that increased patient involvement in rehabilitation training, characterized by active participation an‎d minimal external intervention can improve rehabilitation outcomes by inducing neural plasticity. To achieve this, assist-as-needed (AAN) control strategies have been developed, where the robot controller provides the minimum required assistance to the patient. Given the growing application of robotics in rehabilitation an‎d their crucial role in facilitating recovery, in this research novel control algorithms are proposed to enable intelligent assist-as-needed behavior in upper-limb rehabilitation robots. In these systems, in addition to uncertainties in the robot’s dynamics, the inclusion of the human as part of the closed-loop system—with unknown dynamics an‎d variable interactive inputs—introduces further challenges to the control design. Accordingly, the proposed methods adapt robot assistance based on real-time identification an‎d eva‎luation of human performance, ensuring that the robot intervenes only to the extent necessary. Three distinct control approaches are developed to implement this adaptive behavior. In the first approach, robotic rehabilitation is modeled as a two-player non-zero-sum game, leading to the development of an adaptive optimal control strategy for achieving AAN behavior. The performance of the proposed approach is illustrated in several simulations an‎d experimental studies, using a 7-DOF Franka Emika robot arm. In the second an‎d third approaches, the adaptive neural network algorithm is employed to implement a variable admittance model for AAN control. In these methods, a novel performance index based on the interactive power input is proposed to identify the human’s sharing role in real time, treating the human as an active contributor to the system’s dynamic energy flow. The primary distinction between the second an‎d third approaches lies in their respective strategies for addressing the safety of the interaction. In the second approach, a modified virtual energy tank method is introduced to incorporate the passivity preserving conditions an‎d guarantee safe human–robot interaction. In contrast, the third approach adopts a Lyapunov-based adaptive neural network algorithm to ensure the stability of the interactive system. In this approach, the adaptation laws are specifically designed to guarantee both the system’s stability an‎d the boundedness of the neural network weights under variable interaction conditions. The second an‎d third approaches are validated experimentally using a 7-DOF UFactory xArm7 robot. While the focus of this study is on end-effector-based robots for upper-limb rehabilitation, the proposed control methodologies can be generalized to a wide range of rehabilitation robotic systems. Collectively, the results demonstrate that the proposed approaches achieve two main rehabilitation objectives: encouraging active an‎d safe human participation, an‎d enhancing therapy by providing personalized assistance tailored to individual abilities an‎d conditions, thereby reducing the need for therapist intervention.

مهدي كشميري , حميد صادقيان

سعيد بهبهاني

محمد دانش , مرضيه مجدراصيل , عليرضا ميرباقري