Learning basic unit movements for humanoid arm motion control

Abstract

Manipulation skill is important for humanoid robots to live and work with humans, and arm motion control is essential for the manipulation accomplishment. In our research, we hope our robot execute a manipulation task by combining basic unit movements (BUMs), thus making the manipulation easier and more robust. So in this paper, we firstly define BUMs which actually can be regarded as basic components of any arm motion. Then we propose a learning approach for the robot to execute BUMs, which means knowing the current state, the robot learns how to move his arm to accomplish the given BUM. Considering the complexity and inaccuracy problems in solving the inverse kinematics, the proposed approach is basically building an internal inverse model and the robot directly learns in the motor space without any inverse kinematics. Taking advantages of the powerful capacity of Deep Neural Networks (DNN) in extracting inherent features, the auto-encoder is employed to formalize our model. Experimental results on MATLAB simulation as well as PKU-HR5II humanoid robot reveal the effectiveness of the proposed approach. The robot can successfully execute almost all the BUMs in the whole workspace of his right arm with the accuracy of 98.49%.