Learning to reach object with the desired pose by using visual information

We focus on the object reaching problem for eye-on-hand robot system: how to control the end-effector to reach the exact position and orientation according to the visual information of object. The exact position and orientation are defined as “reaching pose” in this paper. Two issues are mainly concerned and this leads to two contributions in this paper: (1) Find the relationship between reaching pose and object pose in the image. Unlike traditional methods based on complex calibration, supervised learning algorithm is introduced to learn the relationship based on training data which are collected automatically. This method has two main advantages: 1) it can be implemented with little human assistance and 2) it can handle non-linear case. (2) Represent the object pose in the image. A novel and robust algorithm is proposed to extract the feature points. The feature points are robust and invariant under translation, rotation and scaling. Then those feature points are sorted into a sequence which corresponds to the input-variable vector. The sequence is also robust to make sure the consistency of the meanings of the input variables. Finally, the experiments show the validity of our method.