Curriculum-based Humanoid Robot Identification using Large-scale Human Motion Database

概要

Identification of an accurate dynamics model remains challenging for a humanoid robot. It requires to prepare a good initial model and to solve a complex optimization problem for sampling a wide variety of motion data.

To address these problems, I propose a curriculum-based identification approach in this paper. The curriculum guides the sampling process of the motion data so that the accurate dynamics model is gradually learned from an unreliable initial model. Therefore, the good initial model is not required in my proposed method. Moreover, I avoid to solve the complex optimization problem by creating the curriculum using a large-scale human motion database.

I evaluated my proposed method in a simulation experiment and demonstrated that my curriculum successfully guided to obtain a wide variety of motion data. Consequently, an accurate model of a 18 DoF upper-body humanoid robot could be identified with my proposed method.

Moreover, in the end of this thesis, I introduced the possible approach for improving the modeling accuracy by using non-parametric learning schemes. Through the derivation of the stochastic dynamics model, it was shown that the method could be reduce the errors which was not able to be compensated by fixed parameters.