AbstractsComputer Science

This dissertation investigates the efficient imitation learning framework based on principal component analysis to generate desired and abundant motions using few human demonstrations in the workspace. This framework is inspired from human motor learning in upper body. In real life, a robot is expected to autonomously act with varied and abundant motions. Imitation learning might be an efficient approach to enable a robot to generate natural and abundant motions. The framework comprises off-line and on-line processes. In the off-line process, human demonstrations are used to develop a motion database. The database covers the workspace and includes robot properties. The evolved database has a clustered structure for efficiency. In the on-line process, a robot can generate desired motions using a real-time motion reconstruction method based on PCA in real time. However, during the motion reconstruction by few, slight movement, or noisy, it occurs the over-response problem. To prevent that, we uses Regularization method, and curve fitting method to reduce the tolerance. The performance of this method is verified through two case studies. The proposed framework is applied to the generation of reaching motions to an object on a table and a shelf.