並木 明夫

<p>In recent years, researches on manipulation systems that can respond quickly to changes in the environment are increasing. Our research group has studied a juggling robot as one example of them. If the robot is a multi-joint manipulator and moves at high speed, the influence of the dynamics of the robot can not be ignored. Our previous studies have mainly used simple PD control based on the difference between the target joint angle and the actual angle. However, due to the influence of dynamics, tracking errors remain at high speed operation. In order to solve this problem, in this paper nonlinear model predictive control (NMPC) is applied to a high speed robot. Since NMPC has a large calculation load, it applies only to axes with large tracking error. Experimental results are shown to verify the proposed method.</p>

In robotic manipulation visual servoing is important to achieve dexterous and accurate handling. And one of the most important problems is occlusion by the robot's body. It is difficult to know the position and orientation of the object being manipulated when the robot manipulator handles it because it is hidden by the robot itself. If its position and orientation are estimated considering the occlusion by the robot, the robot can manipulate it by visual servoing more dextrously. In this paper, we propose an algorithm to estimate the position and orientation of the manipulated object during manipulation by using a 3D sensor. The 3D information of the total environment is observed by the 3D sensor, and the observed 3D point cloud information is classified into the manipulator, the object, and the others based on their 3D model. The proposed algorithm is verified in an actual robot manipulation system.

The demand of robots, which is to work under the hazardous environment scene instead of human, is getting increased in recent years. Such robots should be controlled directly by human operators to make flexible decisions under such conditions, so our study is to develop the system of using master-slave with high maneuverable master device. In this research we propose a new method of master-slave operation with manipulating multiple numbers of slave robots by single master operator. In particular, the method is to transfer the coordinate system of operator's hand directly to the manipulated object, but not to robot's hand. It is efficient in some situations which is found difficulty with using only one slave robot, such that handling a large object. Moreover, we aimed to realize the proposed system without a feeling of wrongness, by evaluating the movability of the slave arm, using Gaussian Mixture Model of the hand's movable area.

Juggling is one of tasks involving skillful dynamic manipulation, and visual information is essential to perform juggling. We achieved a two-ball juggling task with our high-speed robotic hand-arm system which had three general-purpose fingers and high-speed vision system. But it could not afford enough long operation time for juggling because the trajectory was not optimized. Therefore this algorithm often caused miss of catching the ball. In this study, we work on the improvement of the robot's trajectory at the juggling. We get the 3-dimensional hand position in human ball juggling by a 3D sensor. As a result the characteristics of human juggling is obtained. Based on the motion analysis, we make the throwing motion whose orbit is a circle including internal and external motion. We are successful in the throwing and catching motion by a new program.

In this study, we propose a novel air-hockey robot that can change its actions according to the behaviors of opposite human player. In the robot, its attack behaviors are optimized during an air hockey game by recognizing the motions of the puck and the human hand. First, a method of how to calculate the attack position is explained. Secondly, the value of an attack calculated by the information of the puck and mallet including the position of human hand is explained. Then, we show the data of experiments and explain the reason of robot behaviors by the calculated value.

In this paper, we propose a sword-fighting robot system controlled by a stereo high-speed vision system as an example of human-robot dynamic interaction systems. The robot recognizes motions of an opponent human player quickly by using CAMshift. CAMshift is an tracking algorithm based on the pattern of the color histogram of a target, and it is robust against disturbance of image noise. The vision system recognize the positions of both the hand of a human player and the sword grasped by the hand. We suppose the sword is controlled by a 6DOF virtual arm and predict its state by using Extended Kalman Filter. The robot achieve defensive motions based on this prediction.

This paper proposes a robot which can pour liquid from bottle. The robot consists of a 7-axis robot arm and 4-axis robot hand, a high-speed vision, and the robot is controlled based on visual information. The robot system has the abilities to pour liquid smoothly and exactly, is independent of target. A system designer can easily adjust for target by setting several specified parameters. In this paper, first, pour motion design for pour exactly is proposed. This motion plan consists of 3 step which are extracted from human motion. Secondly, vision processing to get liquid surface information. Thirdly, drop position control using vision feedback. Finally, the data of experiments are shown and the effectiveness of the propose method is verified.

In recent years, various robot hands and arms have been developed for achieving dexterous manipulation tasks. However there are few robots that cannot only move quickly but also handle tools dexterously. Kendama motion is one example of dynamic manipulation and skillful handling. There are several previous researches about robotic kendama. However, in these researches robots can not use their robotic hands effectively. The purpose of this study is to achieve kendama motion by the estimation of the grasping object based on high-speed visions and CoP tactile sensors. So, we propose a new revised method to estimate the grasping state based on an extended Kalman filter. As a result, the robot has succeeded in catching motion of kendama.

In recent years, the demand for robots to perform various tasks in dangerous environments has increased. Operation type robots are more suitable for dangerous environments than autonomous robots because human operators can give suitable judgement. We have developed a new master-slave robot system which consists of both a light weight master system and a high-power slave robot, and it is useful for quick and powerful remote operation. However, in quick motions of master-slave systems, self-collision avoidance become a serious problem. In this paper, we propose a novel self-collision avoidance method for master-slave robot systems. In the proposed method, smooth transitions between normal operation and self-collision avoidance are performed. In the self-collision avoidance, the robot are controlled so as to be positioned on the nearest point within its operation permitted area. We show that the proposed method is effective for our master-slave robot system in experiments.

Juggling is one of tasks involving skillful dynamic manipulation, and visual information is essential to perform juggling. We have already achieved a two-ball juggling task with our high-speed robotic hand-arm system which has three general-purpose fingers and high-speed vision system. This is the world's first attempt at achieving two-ball juggling in one hand using a robotic hand-arm. However, juggling motions become sometimes unstable. One of main problems is how to determine the catching point. In order to achieve stable juggling motions, we propose a new method to determine it by considering manipulability at the catching point. In this paper, we explain the method, and we show the results of the numerical simulations.

This paper proposes a teaching system for multi-fingered robot hands based on kinetic information of grasped objects. The goal of this system is to teach manipulation motions of multi-fingered robot hands intuitively and effectively. This paper proposes an algorithm of the trajectory generation of robot hands. An appropriate trajectory of robot hands is calculated according to the desired kinetic trajectory of a manipulated object under pure rolling contact constraints. In this paper, we show a revised algorithm of the trajectory generation and experimental results of the teaching system to verify the effectiveness of them. Also, future works to be improved are discussed.

Most of robots has link mechanisms consisting of several links and joints, and the joint angle sensors are used to measure the poses. If any one of joint angle sensors breaks down, the robot cannot know its correct pose and can not move correctly. And, if there is a gap between the true value and the sensor value of the joint angle, the robot can not also know its pose. So, if the joint angles and the pose can be estimated from a camera image, the robot can solve the above problems and be controlled without joint angle sensors. In this paper, we propose an estimation method of robot manipulator's poses by only a RGB-D sensor without joint angle sensors. In the proposed method, the 3D model of the robot is compared with the acquired RGB-D images, and the joint angles are estimated by using the depth information and the color information. The proposed method was verified by using a multifingered robot hand.