| Title |
Effective Supervised Learning for Robotic Grasp based on Mixture Density Networks |
| Authors |
김병완(Byung Wan Kim) ; 박영빈(Youngbin Park) ; 서일홍(Il Hong Suh) |
| DOI |
https://doi.org/10.5573/ieie.2020.57.6.52 |
| Keywords |
Deep Learning; Supervised Learning; Robotic Grasp |
| Abstract |
In this paper we introduce the effective supervised learning methods and robotic grasping system that performs robot grasping by learning deep neural networks with a small amount of dataset. The process of extracting feature information of an object grasp and estimating the joint angles of the robot are performed at one time from one deep neural network. There is a time-consuming and difficult pre-processing for deep learning in the robot grasping that sufficient data such as a high-quality three-dimensional object model composed of multi-modal sensor information must be secured. We use a relatively small amount of RGB images captured by the head camera and wrist camera to minimize the effort. We present a system that performs the entire grasping operation with two identical structures of deep neural networks, and generate an image focused on a target object using a semantic segmentation technique to maximize efficiency by using it as an input of the first deep neural network. We propose a method to solve various one-to-many matching situations that occur in robot grasping by modifying the Mixture Density Network(MDN) to output the amount of change from the current to the next joint angles. Each learning weight of joints is trained as a learning variable or random constants are given to the training weights of robot joints in the kernel function. Training systems show a high improvement of the success rate. In the experiment of grasping the new pose of the known objects used for learning, the grasping success rate was 81%, and 80% in the experiment of the alternative objects not used in learning. |