A Method to Make a Robot Understand What was a Target Object in Motion Copying System

This paper presents a novel method to identify a target object based on position and force data during motion demonstration. MCS is a system that copy and reproduce a skillful human motion through bilateral teleoperation. Even though, MSC can teach a robot how to move, a robot cannot recognize a target object because conventional MCS does not record environmental information. In proposed system, a camera is used to add environmental information. We use object detection algorithm to detect not a target object but a robot itself. The detected robot area is used to combine manipulator's information in image space and in robotic work space. By checking detected robot area and haptic information, we can obtain a region around a target object automatically. After automatic target image data collection, we train Convolutional Auto Encoder(CAE) so that CAE can extract target information. The proposed neural network can selectively detect the target object for MCS, which means a robot understand a target for MCS. The results of end effectors' detection and target object extraction are shown in images through experiments.

I. Introduction

Robotic automation is expected to replace human work force. Although most of products such as cellular phone, electrical circuit or PC are made by robots, there are still many tasks that robots cannot do such as cooking, crops harvesting and so on. In order to do a task by a robot, a robot has to decide its motion. There two types of motion. One is position. The other is force. There are two challenging tasks for robotic automation. One is that a robot has to sense and control its reaction force. Force control, impedance control and hybrid control is a control to realize a flexible motion by a robot [1] [2]. The other is that a robot has to plan acting force and trajectory(position) based on target states through time. These tasks are difficult for a robot. However, these tasks are relatively easy for a human. Therefore, a lot of researchers focus on how to teach human motion to a robot. Motion copying system(MCS) is a motion teaching method. The novelty of MCS is that it teaches force information through bilateral teleoperation which transmits haptic sensation during teleoperation. One major problem according to MCS is that if a target or environment is changed between motion recording and motion reproducing phase, the motion reproduction cannot be realized. There are several compensation methods for MCS. Instead of realizing position tracking, yajima and miura tried to realize velocity or acceleration command tracking so that motion reproducing phase is not affected by position fluctuation from object's original position [3] [4]. Nozaki proposed a compensation method which change time scaling in motion reproducing phase [5]. If object is located farther, it extends time scale so as to make contact time achieves a desired time. Nagatsu define the difference in motion between motion saving and reproducing phases as a Environmental disturbance and proposed a method to compensate such disturbances [6]. These kind of compensations make MCS more adaptable to changed environment. However, these methods can only realized under the condition that a robot contacts with a target. Then, the problem will be how to make a robot touch a target automatically. In order to touch a target, a robot needs to know what is a target for a task. Although, sun used a stereo camera to sense position of a target. His recognition system is not based on recorded motion. Therefore, we propose a method which makes a robot understand what is a target for MCS in this research. The idea to make a robot understand MCS's target is a new concept. Although there several researches that analyze recorded motion in MCS. These motion analysis is based on a robot itself. These are not based on target environments [7]–[9].