The physiological responses that arise from human-robot interaction may vary across subjects in magnitude and rate. Such individual variations may require instance-based learning over model-based learning algorithms. For instance, a wearable assist-as-needed exoskeleton may require real-time progress data to provide the appropriate level of support to a specific user. In this study, an instance-based learning algorithm was developed and integrated with a computed torque control law. Real-time bio-signals, in the form of electromyography (EMG), were tracked during a predetermined time window to quantify an adaptive threshold value and to control the torque at the exoskeleton joints. These signals were fed to the algorithm, which instantly learned and determined the support needed to accomplish a desired task. The algorithm was tested on a 5-degree-of-freedom wearable exoskeleton used in the automation of upper-limb therapeutic exercises. Results indicated that the algorithm offered the ability to adjust assist-as-needed support instantly based on the amount of muscle engagement present in the combined motion of the human-exoskeleton system.