Generative Adversarial Network Based Human Movement Distribution Learning for Cable-Driven Rehabilitation Robot

Movement distribution analysis can reveal the body’s changes from training with rehabilitation robotic assistance, and the distribution result has been used to develop robot control scheme. However, movement distribution modeling and further validation of the control scheme remain a problem. In this study, we propose a generative adversarial network (GAN) to learn the distribution of human movement, which will be used to design the control scheme for a cable-driven robot later. We preliminary collect a movement dataset of ten healthy subjects following a circular training trajectory, and develop a GAN model based on WGAN-GP to learn the distribution of the dataset. The distribution of the generated data is close to that of the real dataset (Kullback-Leibler divergence = 0.172). Ergodicity is also used to measure the movement trajectories generated by our GAN model and that of the real dataset, and there is no significant difference (p = 0.342). The results show that the developed GAN model can capture the features of human movement distribution effectively. Future work will focus on conducting further experiments based on the proposed control scheme, integrating human movement distribution into the control of real cable-driven robot, recruiting subjects for robot training experiments and evaluation.