The main objective of this research was to introduce a smart assist-as-needed control system that helps elderly or partially paralyzed individuals. To ensure that a smart and compliant controller, in each cycle of the gait is developed, we adapted the target impedance gains and feed-forward force of the assistive mechanism according to a learning law. A strength metric was defined to determine when the human needs assistance. Then, a cost function was introduced and the gains are modified to reduce the cost function. Applying the proposed controller, the interaction force between patient's limb and robot was reduced in cases wherein user has sufficient strength for task execution and minimizes the position error in the sub phases where users need assistance. Using the Lyapunov stability theorem, the stability of the closed loop system was proved without and with parameter uncertainty, respectively. The results of simulations and an experiment on an exoskeleton showed that the proposed adaptive impedance control improves performance to a level substantially higher than that achieved with constant impedance control. © 2018 Elsevier Ltd