Gitae Kang

Exploration and reconstruction of unknown object by active touch of robot hand

This paper proposes a method of exploring the global shape of an unknown object using information on local geometric features. In the first, we introduce a rolling and sliding motion of a fingertip with a force/torque sensor to estimate an unknown local curvature. Also, a recognition algorithm for local geometry using normal curvature equations is presented, which are composed of principal curvatures and principal direction. Finally, to reconstruct the global shape of the object, we propose an interpolation method using principal curvatures at contact points. The proposed method is verified using a hand-arm system consisting of an industrial robot arm and an anthropomorphic robot hand with a 6-axis force/torque sensor. The effectiveness of the proposed method is experimentally validated for different type of objects.

Design of anthropomorphic robot hand with IMC joints

This paper presents a novel anthropomorphic robot hand with IMC joints. IMC joints are applied to robot hand to grasp various objects regardless of the size and get large overlapped workspace between thumb and other fingers, which allow all the fingertips to be located on a single spot in a wide range. Desinged IMC joints of robot hand are always rotated with 1:1 ratio by mechanical constraint using passive tendon. Actuation module consists of miniature BLDC motor and ballscrew drives two IMC joints with a single motor. To evaluate the advantages of robot hand with IMC joints, grasping experiments are performed.

Design of backdrivable soft robotic finger mechanism

This paper presents a new joint actuation mechanism, called Active DIP-PIP (ADP) joint, for robotic finger. The mechanism consists of a pair of moveable pulleys and springs to generate both linked and adjustable motion. While the set of DIP (Distal-Interphalangeal) and PIP (Proximal-Interphalangeal) joint shows coupled movement in free space, it moves adaptively when it contacts with an object. The torsion springs attached on each joints ensure 1:1 ratio movement while finger is moving freely and produce additional joint torque while grasping an object. In addition, actuation module composed of miniature BLDC motor and ball screw allows each joint to be back drivable.

Kinematic design optimization of improved branched tendon mechanism using genetic algorithm

This paper presents the improved branched tendon mechanism by including additional design parameters to the original branched tendon design, and a kinematic design optimization technique for this mechanism using genetic algorithm. The significance of additional design parameters and the feature of improved branched tendon mechanism are also explained. Unlike traditional joint pulley-tendon mechanism that always has the same length of moment arm, the improved branched tendon mechanism uses special tendon which is divided into two just before it is attached to the remote link. By optimizing these divided two different lengthes of tendons and other design parameters, creating various moment arm on a single joint with respect to the flexion angle of joint and limiting maximum moment arm to ensure all the tendons to be inside of the fingers outer frame are possible at the same time. Total 12 variables, 4 given and 8 independent, are used to represent the mechanism mathematically and the objective function is defined to maximize the moment arm throughout the flexion of the joint while not exceeding joint radius. Optimizing the kinematic model of improved branched tendon mechanism with genetic algorithm, it is possible to minimize the loss of the moment arm to 311% throughout the flexion. Meanwhile, overall actuation mechanism becomes much simple than traditional joint pulley-tendon mechanism.

Force/torque sensor calibration method by using deep-learning

The force/torque sensor is an important tool that gives a robot an ability to interact with their usage environments. Calibration is essential for these force/torque sensors to convert the raw sensor values to accurate forces and torques. However, in practice, the multi-axis force/torque sensor requires complex multi-step data processing, because of the coupling effects and nonlinearity of sensors. Moreover, accuracy is not guaranteed. To solve this problem, we propose an accurate force/torque sensor calibration method that can calibrate the sensor in single step by using deep-learning algorithm, and introduce the method for modeling the DNN(deep neural network) used in this calibration process. In addition, we also explain some tricks for learning, and then verify the calibration results through several experiments.

Sampling-based path planning with goal oriented sampling

Path planning in complicated environments is a time consuming and computationally expensive task. Especially in high-dimensional configuration spaces with complex obstacles, searching for a proper path while avoiding collisions is still challenging. This paper presents an improved sampling-based algorithm, called the Goal Oriented sampling method (GO sampling) that quickly generates an initial solution overcoming these problems. GO sampling extends the sampling method of the Rapidly-exploring Random Tree (RRT) algorithm. GO sampling is able to identify the initial solution in a shorter time than that of the RRT algorithm and shows significant improvement in computational efficiency. The algorithm is evaluated with simulations in 2D and 3D space.