Motion Generation in Hybrid Control Converted from Human Motion Data

Robots are desired to replace human labor. One method for having robots perform human tasks is imitation learning. However, obtaining a large amount of human motion data necessary for imitation learning is difficult due to physical constraints. Therefore, to learn motions from a small amount of data, a method to extract features of motions from the motion data and to obtain a generic representation of the motions is being considered. In addition, although there are methods that consider position and force in imitation learning, the control stiffness is fixed. This paper determines the control directions of position and force for one motion data by Principal Component Analysis (PCA), and the control stiffness is determined. We also propose a method to generate a motion suitable for each state by dividing the motion into multiple states using the results of PCA and estimating the states during motion generation. Simulation is performed to verify the results of motion generation based on the proposed representation.