This paper proposes a multi-joint-gripper that achieves envelope grasping for unknown shape objects. The proposed mechanism is based on a chain of Differential Gear Systems (DGS) controlled by only one motor. It also equips a Variable Stiffness Mechanism (VSM) that controls joint stiffness to relieve interferences suffered from external environment and to achieve a dexterous grasping. The experiments elucidated that the developed gripper achieves envelop grasping; the posture of the gripper automatically fits the shape of the object with no sensory feedback. And they also showed that the VSM effectively works to relieve external interferences. This paper shows the mechanism and experimental results of the second test machine that was developed inheriting the idea of DGS used in the first test machine but has a completely altered VSM.
The ultrasonic motor is an actuator that has advantages such as low speed, high torque (good response time), and high positioning accuracy. In this paper, we analitically compare the responsiveness of the ultrasonic motors with the electromagnetic motors using those exercise equation. From the simulation results, a good responsiveness of the ultrasonic motors is clarified in step and frequency responses, when the output torque of both motors are same and a load is smaller than the max angular acceleration of both motors becomes same. Another focus of this paper is that a high-speed camera with high-power lens visualizes how the ellitilcal motion changes. By using the high speed microscope, we examine the relationship between the change of the elliptical motion and the motor responsiveness analytically and experimentally.
This paper presents a design method of upper limb, for human mimtic musculoskeletal humanoid Kenshiro. The design of his upper limb is realizing detail features of muscles, bones and the adhesive relation of the two. Human mimetic design is realized by focusing on the fact that joints are being stabled by muscles winding around the bones. On designing a human mimetic robot, we define criteria on what it is to be “human mimetic” in a humanoid robot. Focusing on adhesion of muscles and bones, and by following human mimetic criteria, Kenshiro's arm was made. The use of this robot can be expetcted as a human body simulator, by measuring sensor data which can correspond to biological data.
When a robot grasps an object, posture errors between the hand and the object has a major influence on grasp success. So we have proposed to correct it before touching with Net-Structure Proximity Sensor (NSPS). This sensor can detect posture errors and distance to the object within tens of millimeters from the sensor. The sensor has characteristics of fast response (<1 [ms]), simple wiring (only 6 wires), and easy implementation. In this paper, we demonstrate simultaneous control of posture errors and distance of the fingertip before grasping. In order to archive this motion, we use a fingertip equipped with the proximity detector array, which can detect fingertip-object distance and postures simultaneously. Through this preparatory motion (called “pre-grasp”), it is possible to adjust fingertip position to vertical from the object surface. The robot hand can adjust fingertip postures for objects of various shapes automatically. So number of programs corresponding to shapes are not necessary. And it is also suitable for grasping of a fragile object, because it is kept low pressure of the object surface.
A method to detect hazards in the sense of traversability at Tsukuba special zones for robotic mobility (TRM) is presented. Target robotic transporters are supposed to be equipped with a range image sensor that can recognize front road surfaces of pedestrian areas in TRM. The hazards on the surfaces are defined by subjective judgment of transporter users; the users should provide a traversable/untraversable label with respect to each single range image at a TRM location such that those labels reflect surface hazards with regard to roughness, drivability, and positive/negative obstacles. Our approach applies off-line learning to a set of feature vector obtained from those labeled range images and predicts the traversability online. The learning problem is formulated as a two-class as well as a single-class classification; the latter can skip manual labeling required by the former supervised learning formulation. The feature vectors are extracted from both geometric and probabilistic models of range image observations. Experiment results show that two-class classification yields better results than one-class one with regard to the accuracy, nonlinear discrimination by polynomial kernel and Gaussian kernel works in a similar manner, and that performance with 90% true positive rate is achieved in exchange for accepting 20% false positive rate.