Equations (4) and (8) Planned Motions Feedforward Torque ε q∗(t) + - q(t) Figure 2. The learning and control diagram of a BLS-based human-like torque and impedance adaptation. rate of the force and impedance, which is fully different from previous research. Referring to the models from [5]-[13], especially [8] and [12], we construct the motor learning and control diagram for torque and impedance, as displayed in Figure 2. Referring to [12], a sliding tracking error term is built for updating feedforward force. The difference between the proposed framework and other iterative learning methods is that the feedforward force is estimated using feature expressions and their networks, which is similar to the feedforward model based on the radial basis function (RBF) NN in [22]. As introduced in [24] and [8], the joint stiffness matrix increases with torque and muscle activation. Then, the neural feature terms in the feedforward torque block are transformed by adding noise terms to describe the contractile effect of the muscle. The impedance is expressed by the linear combination of enhancement nodes (the idea described by BLS), which is achieved by transformations of feature nodes from the feedforward torque. The impedance changes are realized by increasing the number of enhancement nodes and updating the weights according to the tracking errors. Feedback torque is then achieved based on the