An sEMG-Driven Identification of Human Upper Limb Stiffness and Its Application in Variable Impedance Control

In this paper, we try to estimate human muscle stiffness in a static estimation frame and transfer it to Baxter by space conversion to realize a variable impedance control. In this frame, a neural network is designed to estimating the end force, which is mapped with human upper limb stiffness by the TVSEM model. In this way, human stiffness can be estimated at a specific posture with different force export. Then, the stiffness will be transferred to the robot control strategy and realize posture-limited soft control. For the sake of estimating endpoint force of human upper limb by sEMG efficiently, we test two classical neural network models, Dilated convolutional neural network(DCNN) and Squeeze-and-Excitation neural network(SEnet). To improve the estimated capability, we design a new compound model named Dilated-Convolution Squeeze-and-Excitation neural network(DC-SE). We compared them in several different situations, which contain estimating different quantities of category or estimating the same categories but use shorter or longer sample lengths. To confirm the efficiency of this muscle stiffness estimation frame, we apply it in a variable impedance control experiment. Human upper limb stiffness is estimated and transferred from stiffness space to joint space by space conversion for impedance control needs. And the result of the experiment proves that the method could estimate joint stiffness reliably.