An electromyography-driven central pattern generator model for robotic control application

Central pattern generator (CPG) models have been designed at abstract levels of the rhythm phenomena and widely applied in robot control. The robot controlled with a CPG model is hard to perceive and respond to the motion intention coming from human beings. A new CPG model driven by surface electromyography (sEMG) was presented in this paper to manipulate robots more favorably without loss of their autonomous capability. The CPG model was designed based on an echo state network (ESN) which was a large, random, recurrent neural network. The frequency modulating from inputs to outputs was researched in this study. It was illustrated that ESNs could learn and generalize the frequency transition pattern. The flexion and extension motion of forearms and the sEMG at biceps and triceps muscles were sampled as the teacher signals to train the CPG model. The prediction error of the trained model was analyzed carefully and the model output was applied to control a rehabilitation exoskeleton. Finally, the future work was discussed on the model structure optimizing.