Figure 10(a) shows the path tracking curve of the RRDS in the experiment. It can be observed that the motion path of the robot fluctuated slightly at the beginning. However, it met the allowable error range of the human-robot system. As the number of learning times increased, the tracking error gradually decreased, and the RRDS finally achieved stable tracking for both passive and active direct switching training. Figure 10(b)-(d) shows the velocity tracking curves in the passive and active switching training process. From these figures, it can be observed that the RRDS velocity constraint range of each axis satisfied () vt 1m/s, vt 1 m/s, x # () i # y # () and vt .025rad/s in the passive training stage. With the improvement of the walking ability, the RRDS directly switched to the active training stage at t 8sf = Moreover, . the velocity was constantly adjusted to satisfy the decision errors on each axis within the range ofhttps://www.hocoma.com/solutions/gait-balance/ https://www.cyberdyne.jp/english/products/HAL/index.html