Wavelet based OFTM for learning stirring food by imitation

Abstract

In this research a new robot learning method based on imitation is introduced which enables a robot to learn new trajectories by only one demonstration. This one-shot learning approach is based on Orthogonal basis Functions and Template Matching (OFTM) which was previously introduced by our group and implemented using the Fourier basis functions. In this paper the W-OFTM method is presented which employs the wavelet transform in the OFTM approach. In W-OFTM the wavelet orthogonal basis functions are included in the dictionary of primitive motions, alongside a few well-established templates. One of the major advantages of this approach is enabling the robot to reproduce all trajectories in its workspace. In this research, a thresholding parameter was automatically set in the F-OFTM and W-OFTM methods in order to filter out unimportant coefficients and reduce the occupied memory space while holding the increased error below a certain acceptable value. In the experimental trial, the proposed method was applied to a chef robot in order to learn the task of stirring food. Results indicate that in comparison to the GMM-GMR method, the W-OFTM method provides more accurate results with much less delay. Furthermore, the advantage of the proposed method over the state of the art method increases as the numbers of samples contained in a trajectory increases.