In the musculoskeletal humanoid trying to make use of advantage of the human body by mimicking its structure, ligaments, which constrain the joint movement softly and form the complex range of motion, are no less important traction structure between skeleton than muscles which actuate the body. However, the conventional linear traction structure with wire cannot maintain stable tension depending on the posture change of the joint. In this study, we a propose planar traction structure as traction structure between skeleton, and describe and evaluate its advantages. Also, we attached the planar collateral ligaments to the knee joint, and realized screw-home movement like the human knee joint.

In this paper, a humanoid robot, named Q-bot is proposed, which is used in convenience stores in the future. The Q-bot is designed as a small sized crane based on universal vacuum gripper of its foot sole, which holds its body by vacuum force as a compensation of its small weight. Moreover, the foot vacuum prevents Q-bot from its fall over by reaction force of works. such as driling a wall. The robot has several functions of dual-armed manipulation and omni-directional movability, and the tips of arms and a foot equipped with a universal vacuum gripper for adhering to uneven surface. This paper reports the development status of the robot and the experiment which is possible to prevent the robot from falling when the robot adhering to the ground, grips heavy objects.

A new design of bipedal walking robot that utilizes 6-dof parallel link mechanism for each leg is proposed. Reducing the weight of the legs is a crucial requirement for realizing rapid walking and fall avoidance involving multiple stepping. The parallel link design enables embedding all leg actuators in the torso and thereby significantly reduce the weight of the legs. The result of basic kinematic analysis of the proposed leg mechanism is reported.

A novel lower leg mechanism for humanoid robots is designed. The pronation/supination motion of the ankle is added to an existing planar complex knee-ankle mechanism that enables firm support against gravity and natural swinging by gravity. The increase of total moment of inertia is prevented by a linkage mechanism that remotely rotates the ankle, where the motors are assigned at the knee part. A prototype of the machanism was designed on 3D CAD.

To realize a robust automobile driving behavior of musculoskeletal tendon-driven humanoids, we developed a new six-axis force measurement module with a core-shell structure. The core-shell structure consists of core part, shell part and loadcells between two parts, and a six-axis force sensor with this structure enables space saving, high load capacity and freedom of designing of sensing surface at the same time. By developing a foot unit incorporating two core-shell structural six-axis force sensors on its toe and heel, we realized behaviors of stepping on a brake pedal and recovering from slipping with a lifesized musculoskeletal humanoid ”Musashi” and confirmed the effectiveness of the core-shell structure.

In recent years, researches on musculoskeletal humanoids are on the march. However, there are some challenges such as unmeasurable transformation of body structure and muscle path, and difficulty in measuring own motion because of lack of joint angle sensor. In this study, we suggest two motion acquisition methods. One is a method to acquire antagonistic relations of muscles, and the other is a method to acquire correct hand trajectory by visual feedback. Finally, we realize badminton shuttle-hitting motion of “Kengoro” by applying two acquisition methods.

The recognition of environment and self is important in researches of musculoskeletal humanoids and there are various sensory information they can utilize. Among them, sense of hearing is useful to recognize the result of own motion involving tapping sound. Human can improve own motion to make intended sound like playing a percussion instrument. We aimed that human mimetic musculoskeletal humanoid acquire the motion based on tapping sound. In this study, we developed sound recognition system to discriminate slightly different tapping sound. Moreover, we developed motion acquisition system based on the result of tapping sound recognition. Through these approaches, we succeeeded in acquiring a motion to make an intended snare drum sound with musculoskeletal humanoid “Kengoro”.

Daily-life humanoid robots are expected to perform a variety of tasks in the household environment. Unlike the tasks in factories, which are probably unique in each robot, there are enormous types of works desired to be done, which can also vary from person to person. In such a situation, those robots should have the ability that users can teach them new motions easily. In this paper, we introduce a new interactive-teaching system utilizing human demonstration and dialog. Each motion is labeled with the motion name which is mentioned while teaching the motion. It enables easily requesting the robot to perform a certain action, and also associating with the condition where the task should be done.

Self collided posture can reduce the load of joints because contact points between colliding links can support the mass of body. To realize 3D multicontact posture with self collision, there are difficulties in handling joint temperature constraints, contact wrench constraints, and degenerated degree of freedom. In this paper, we propose a method of quadratic optimization control integrating wrench distribution and inverse kinematics in which internal wrench is controlled only in controllable directions to reduce excess internal wrench, to satisfy contact wrench constraints and to keep collided posture with servo gain of joints between colliding links zero. With our methods, HRP2-JSKNTS could pick up an object under a desk with supporting the body with the right arm and squatting with the back of the upper leg on the back of the lower leg.

When a humanoid robot does a task that requires the whole-body fast motion such as tennis forehand swing, it is difficult for the robot to exploit full potential of the actuators’ performance and to modify the joint trajectory in real time. We propose a method of whole-body fast motion planning with balance margin. By using this method, we achieved tennis forehand stroke with a humanoid to see a ball with a stereo camera on the its head.

In this research, we propose a method for observation/action planning for detecting an object by a humanoid robot. We first generate a candidate of observation postures of an object to possibly find an object. The observation posture is determined by introducing the Partially Observable Markov Decision Process (POMDP). To move to the determined observation posture, we use the real time walking pattern generator. The effectiveness of the proposed approach is confirmed by using the humanoid robot HRP-2X.

This paper observes and derives Zero Moment Point (ZMP) of the bipedal robot that equips the trunk embedding the viscoelastic joint and arms. In the physical experiment, ZMP oscillates in the lateral direction largely though it does not oscillate in the anteroposterior direction. The numerical analysis supports the results of the physical experiment, and it indicates that the robot will achieve the stable locomotion by human-like locomotion in which the arm and the other side of the leg swing ahead simultaneously, and each arm swings asymmetrically back and forth.

This paper introduces a control method for master-slave biped robots to walk on steps. Our method is capable of dealing with the contact between a part of foot and a step. In our method, the target center-of-gravity position of the robot is determined by the command from the master devices. In addition, the target position is adjusted according to the attitude of the lifted foot. Our method is tested by a simulation of walking on a step.

This paper proposes a rough terrain walking method for master-slave bipedal robots. This method uses load cells mounted on the foot soles to fit the attitude of the lifted foot to terrain. In addition, the desired trajectory of the center of gravity is generated from the positional relation between the feet to control the position of the center of gravity. The proposed method is validated through a realtime simulation environment involving haptic devices.

Multi-copter is required to move ascending flight, descending flight and hovering flight with adequate controlled behavior in the air environment. This study aims at acquiring autonomous behavior for such a multi-copter by physically modeled simulation. This multi-copter is modeled by use of the physical engine (PhysX) in a virtual environment. We use neuro-evolution to have the multi-copter acquire autonomous behavior. Namely the multi- copter behavior is controlled by an artificial neural network (ANN), The ANN is evolved to let it learn proper behavior by evolutionary computation (GA). Simulation results verify that the multi-copter can acquire autonomous behaviors.

A guide service robot is required to move on any obstacles ground with adequate controlled behavior in indoor environment. This study aims at acquiring autonomous behavior for such a guide service robot by physically modeled simulation. This guide service robot is modeled by use of the physical engine (PhysX) in a virtual environment. We use neuro-evolution to have the guide service robot acquire autonomous behavior. Namely, the guide service robot behavior is controlled by an artificial neural network (ANN). The ANN is evolved to let it learn proper behavior by evolutionary computation (GA). Simulation results verify that the guide service robot can acquire autonomous behaviors.

There were various studies of Humanoid’s bipedal walking. However, most of the researches uses a simplified model. In this research, we study humanoid in hopes of realizing robot walking like a human and consider a gait model of humanoid robot including slipping, bumping, surface-contacting and point-contacting of foot is discussed. First, to stabilize attitude and prevent humanoid from falling down, we proposed Visual-lifting Approach. This method use it’s eyes. Humanoid judged surrounding condition and keep it’s head position high. Next, we need input torque to step forward by feed-forward control but it determined by try and error. This might be waste of energy. So, to solve this problem, I use genetic algorithm(GA). In the report, I explored the best input by GA method and investigated the result.

Biped locomotion created by a controller based on ZMP known as reliable control method looks different from human’s walking on the view point that ZMP-based walking does not include tipping-over state. However, the walking control that does not depend on ZMP is vulnerable to turnover. Therefore, we propose walking stabilizer based on Visual-lifting Approach to enhance standing robustness and prevent the robot from falling down. Simulation results indicate that this stabilizer helps stabilize pose and bipedal walking even though ZMP is not kept inside convex hull of supporting area. Moreover, we point out that arms begin to swing asymmetrically by dynamical coupling among body links without input torques and verify the effects of the arms’ swing on the ice-skate walking.

We use an acceleration sensor with Particle Filter on Episode (PFoE), which is a teach-and-replay method for mobile robots. The PFoE which uses an acceleration sensor is more robust to the noise of motion and sensor information than the PFoE only using distance sensors. With an acceleration sensor, PFoE extends its ability to follow the behavior on the replay phase. Unlike the distance sensors, the acceleration sensor information is available while the robot is acting and it does not become saturated easily. The robot’s behaviors with/without acceleration sensor are compared by a task with an actual robot.

In many mechanical systems, control parameters are tuned by a human expert through trial and error, which is labor-intensive and time-consuming. For example, electronically controlled transmissions (ECT) require such parameter optimization. To address this issue, we propose a parameter optimization system for ECT by using Hierarchical Stochastic Optimization (HSO) that is able to handle multimodal objective function. The optimizer learns better parameters which show high peformance in all metrics and satisfy all constraints. In the experiments, we use multi-physics simulators and optimize the parameters for ECT. Through experiments, we demonstrate that our HSO can identify several modes of the objective function and is more sample-efficient than random search and a human operator.

The musculoskeletal humanoid has many benefits that human beings have, but the modeling of its complex flexible body is difficult. Although we have developed an online acquisition method of the nonlinear relationship between joints and muscles, we could not completely match the actual robot and its self-body image. When realizing a certain task, the direct relationship between the control input and task state needs to be learned. So, we construct a neural network representing the time-series relationship between the control input and task state, and realize the intended task state by applying the network to a real-time control. In this research, we conduct accelerator pedal control experiments as one application, and verify the effectiveness of this study.

In this study, we develop an environment adaptability enhancement system with active information search in real space and high efficiency motion search in virtual space for multi-degree of freedom disaster response robot. For such robots to work safely, estimating the work achievability before executing it is important. However, in disaster sites, failed motions such as falls increase the risk of secondary disasters, so it is difficult to learn motions by trial and error. Thus, the robot obtains the environment information in real space and reproduces them in virtual space, and then performs motion search efficiently in virtual space. Due to the limitation of the reproduction accuracy, for finer adjustment, the robot performs active touches to acquire ground properties such as friction and hardness in real space. The experimental results showed that the proposed system could judge work availability by trying robot motions and provide the robot with efficient motions.

Among the cardiac SPECT examination, in the examination using the 99m-Tc formulation as the radiopharmaceutical tends to accumulate many radiopharmaceuticals in the organs other than the cardiac. As the result, artifacts are generated from organs other than the cardiac during image reconstruction. Artifacts can cause troubles in diagnosis and it needs to be removed, but it takes time and effort. In this study, we believe that if we use Deep Learning to remove non-cardiac accumulation will help eliminate artifacts. Detect the cardiac area by using Semantic Segmentation and construct an auxiliary system for removing artifacts.

It is desirable for a robot to be able to accomplish assigned tasks even in a changing environment. Deep reinforcement learning draws attention recently as a learning method for autonomous mobile robots. This research picks up an example mission where an experimental robot TurtleBot3 Burger reaches the target point avoiding obstacles. Simulation experiments are performed to compare conventional Q-learning and DQN．As a result, DQN is superior to Q-learning in that it can effectively learn the proper actions using the sensor output directly as the input.

Catastrophic forgetting is one of the most challenging problems of (deep) neural networks, but autonomous robots, which would acquire many tasks in real life sequentially, require to resolve or mitigate it. Modular networks are expected to mitigate this problem since it can exploit different modules for respective tasks. However, this approach would waste the learnable parameters due to duplication of common tasks in given tasks. Furthermore, if given tasks, e.g., locomotion control of legged robots, are with high state-action spaces and are difficult to be learned, exploration is required for a long time, which causes the catastrophic forgetting. Hence, this paper proposes the way to divide the locomotion control tasks modularly and hierarchically. To this end, fractality of fractal reservoir computing is utilized so as to transfer the learned knowledge of one leg control to the other legs control.

The parallel link mechanism for robots is expected as a robot component for realizing such as high-speed motion, high-rigidity for a load, high-precision positioning, and so on. One of the parallel link mechanisms is the 3-DOF spherical parallel mechanisms. This is that the Base Plate is connected to the End plate with the three units of 3 joints and 3 links. We consider to use the 3-DOF spherical parallel mechanisms as a joint actuator, the shoulder or wrist of the robots, and so on. In this paper, we propose the inverse kinematics of the 3-DOF spherical parallel mechanisms using a spherical trigonometry and perform motion analysis of the 3-DOF spherical parallel mechanisms.

Parallel link mechanism for robots is expected as a robot component for realizing such as high-speed motion, highrigidity for a load, high-precision positioning, and so on. In its practical usage, however, we often meet a problem of kinematics in which we cannot sometimes define forward and/or inverse kinematics functions. Therefore, we investigate a kinematics problem of parallel link type spherical mechanism where two simple and popular spherical mechanisms are placed to be opposite to each other. In this paper, we present the forward and the inverse kinematics function and their derivation process of the parallel link type spherical mechanism by using spherical trigonometry and the validity of these functions with a numerical experiment.

Customer behaviors represent their interests in the products. In recent years, with the spread of security cameras, it has become possible to record the customer behaviors. However, it takes a great amount of time and effort to analyze the recorded customer behaviors manually. Therefore, automation of customer behavior classification is required. The conventional two-stream I3D composed of 3D convolutional neural network can classify the continuous behaviors of trimmed videos. We extend the two-stream I3D to classify the behaviors from untrimmed videos. Our experimental results show improvement of precision, recall, F-measure, and accuracy using majority decision method for behavior classification.

Medical radiologist diagnoses a medical condition from several varieties of medical image, e.g. MRI images or CT images. Since it is predicted that patients will increase along with the elderly society in Japan, the researches for classifying the abnormal images based on the Deep Learning technique has increased to prevent diagnosis mistakes. Furthermore, there is an increasing need to perform mutual transformation from certain medical images to other modality of images, e.g., from MRI images to CT images. Because different modality of images are taken independently, the patient load becomes big issue to take some medical inspections.

Therefore, in this study, we try to transform from the certain MRI image to the other type of MRI image by using the CycleGAN algorithm known as one of Deep Learning methods. There are two variation of MRI images, i.e., normal weighted images and fat-suppressed weighted images.

In recent decades, a number of researches has been focused on development and control of soft robots. There is one approach that uses a neural network to learn a forward model of a robot and control the robot by extracting equivalent equations from the neural network. However, if the training data obtained from the robot only contains partial dynamic characteristics, the linear state equation extracted from the neural network can be barely controlled. In this research, we propose a method to use singular value decomposition in order to reflect deviation of training data. We confirmed that our proposed method can control a flexible arm by simulation.

This study aims to propose a unified processing model for static information and dynamic information with Spiking Neural Network. We implemented a two-types of a ganglion cell model (P-type, and M-type) to leaky integrate and fire model(leaky-IF) and adopt these model to CNN architecture. Simulated result showed that, P-type model responded to static information (e.g. edges, corners), and M-type model responded to dynamic information (e.g. motion, rotation).

After the Great East Japan Earthquake, “Isoyake” has occurred in Tohoku. A large amount of sea urchin occurred due to the “Isoyake”, algae suffered from damage and suffered severe fishing damage. As a countermeasure for this, we are developing a system to discover sea urchin using small ROV and to remove and recover sea urchin. In this paper, we carried out fundamental research to develop an image recognition system to be mounted on a small ROV. We also constructed a real-time image recognition system for practical application. In image recognition, machine learning was performed using OpenCV. As a result, we were able to recognize sea urchins. Also, we could construct a network for real - time image recognition.

This paper proposes the method that the humanoid with the movable binocular eye-camera unit can acquire the ability to recognize distance to an object by itself. Movable eye-camera has advantages such as quick object tracking or wide field of view. However, it’s difficult to get the distance to an object because of the errors of intrinsic and extrinsic camera parameters. We solve this problem by using the neural network that receives the convergence angle and the object position in binocular images as input and outputs the object position in the real world. In the training phase, the humanoid moves its arm randomly and recognizes hand position in binocular images. At the same time, it collects hand position in the real world calculated by forward kinematics as teaching data. We confirmed the validity of the method by using musculoskeletal humanoid "Musashi" and its movable binocular eye-camera unit.

The aim of this study is to develop efficient and fast running robot and running support device. In this paper, we experimentally demonstrate a stable level running of rimless wheel by traction force of weight. Moreover, we demonstrate Froude number and specific cost of transport of the level running.

In this paper, we aim to the human-like, high-speed running that had energy efficiency based on bouncing rod dynamics. In addition, we aim to understanding for essence of running. In this paper, we compare running of our robot with running of top athletes. And, we demonstrate similarities and differences of these running.

In this study, we aim to the human-like high-speed running that had high energy efficiency based on bouncing rod dynamics. In the previous report, we intervened in the robot’s running by the remote manual operation system with wires, and we realized human-like high speed running on the flat ground. Main intervention points are wire drive by lever operation and cord assist to maintain posture. Minimal intervention by an adaptable human is effective for prompt and bold research, but there is a limit in reproducibility, it was difficult to objectively and quantitatively discuss the control logic. In this paper, we report on fully automated biped running experiment system by motorizing the intervention points.

Passive walking and ZMP walking has long been compared. In the laboratory, we have been experimenting with a passive walking robot. We aim to developing 3D walking robot based on passive dynamical mechanism in recent years. The walking robot has been refined more and more by repeatedly carrying out actual machine experiments, and the tasks are becoming clearer. In this study, we adapted a posture control system in horizontal direction using IMU and servomotor to reduce the ratio of assist by the experimenter. As a result, we improved the balance of the walking robot with this system. In this paper, we report independence improvement of 3D walking on level ground.

In this study, we aim to continuous 3 days passive walking. Even if the passive walking robot walks in the same way, it has good or bad condition, and a fall happens unexpectedly. We think that it is possible to avoid fall risk by adjusting the physical parameters of passive walking and keeping good condition. In this paper, we report the adjustment system of passive walking parameters using deep reinforcement learning.

Towards achieving stable locomotion on slippery ground, a novel crawling-like robot has been proposed in our previous research. In this paper, we modify the original model to improve the performance and further extend it to feeding manipulation system．First, we derive the system dynamics and control method．Second, typical sliding locomotion is generated via numerical simulation．Third, we introduce the feeding manipulation system，and numerically show the feeding process．

This paper presents experiment results of rimless wheel robot in incline mode walking. We developed the Rimless Wheel Walker with Wobbling mass (RW-3). RW-3 achieved the incline mode walking and the wobbling mode walking in the numerical simulation. In incline mode walking, RW-3 walks on level ground using the propulsive effects due to inclining torso posture. In wobbling mode walking, RW-3 walks on level ground using the propulsive effects due to wobbling motion. We show the problems in the incline mode walking through walking experiments.

Passive dynamic walker’s walking is stabilized by using Constraining Sole Shape. In this research, we replace effect of Constraining Sole Shape with the Constraining Pin. We called Constraining Pin the Constraining Foot Mechanism. Passive dynamic walker attached Constraining Foot Mechanism can walk while 2 minutes. However apposite height of Constraining Pin on ramp has been unknown. Passive dynamic walker equipped with gyro sensor to derive apposite height of Constraining Pin. Then we measure angular velocity of passive dynamic walker from gyro sensor. On the receiving the results, we consider apposite height of Constraining Pin. Moreover, we consider apposite height of Constraining Pin from numerical simulation results.

One of the features of Passive Dynamic Walking is adaptive behavior. This adaptive behavior has been confirmed by simulation. In addition, changes in the inclination angle of the walking road surface which is a change in the outside world have been confirmed in actual machine experiments. However, the adaptive behavior to the change in the viscosity of the joint, which is the internal change, has not been confirmed in actual machine experiments. For this reason, we developed a viscous generation mechanism that can change the viscosity of the joint. We conducted an experiment using its viscosity generating mechanism. In this paper, we describe the mechanism of viscous generation, the experimental results using this mechanism, and the installation of viscous generation mechanism in a 4leged passive walking robot.

This study has focused on a unique walking mechanism using synergy effect of the parallel link and rocker function of human walking. The parallel link mechanism is constructed from the musculoskeletal system of the thigh. The bi- articular muscles work as the parallel link component and joint torque transfer. The rocker functions of the lower leg work as generating the shock absorption, driving and kicking force of the parallel link mechanism. The walking mechanism was showed as an experimental mechanical model in this study. This approach of human walking will give unique indication of evaluation and rehabilitation.

This paper proposes a new model of an underactuated rimless wheel (URW) that added a reaction wheel for easily generating a non-slippery stealth walking gait. First, we introduce a 3-DOF URW model that can utilize two control torques, and describe the equation of motion. Second, we develop a controller for achieving strict output-following control of the stance-leg angle and angular momentum constraint control simultaneously. Third, we derive an approximate analytical solution of the target initial states of the upper body and reaction wheel. Furthermore, we propose a simple numerical procedure for identifying the target initial state of the nonlinear model, and investigate the effectiveness through numerical simulations.

Towards adaptation to slippery ground surface for underactuated locomotion robots, the authors proposed the method of stealth walking, which avoids energy loss at landing of swing foot. Moreover, they proposed angular momentum constraint control (AMCC), which constraints the horizontal ground reaction force to zero, to make the robots walk on the frictionless road surface stably. In this paper, we investigate a more severe situation, i.e., stable walking on frictionless descending stairs. First, we introduce a mathematical model of an underactuated rimless wheel with an upper body and a reaction wheel for analysis, and develop the control law for achieving trajectory tracking control of the stance-leg angle and AMCC simultaneously. Second, we investigate the validity of the proposed method and discuss the nonlinearity of zero dynamics through numerical simulations.

In this paper, we propose a method for stable gait generation of a 3-link biped robot with damaged motor on one-side. First, we derive the dynamics and control method. Control torque between the torso and one leg is applied while another is passive. Second, we numerically show the period-2 gait of this semi-passive dynamic walking. Finally, the gait properties are analyzed via parametric study. The results show that by adjusting the target control time for the actuated swing/stand leg separately and appropriately, potential barrier can be overcome successfully.

It was clarified that the walking speed of a combined rimless wheel can be indirectly controlled by using an active wobbling mass via entrainment effect in the previous studies. Analyses on nonlinear properties of it have been conducted based on the theory of phase oscillation. To further investigate the dominance relationship among multiple oscillators in the locomotion systems, different rimless wheels and an active wobbling mass are connected by a rigid beam in this study. First, we describe the mathematical model and control. Second, we discuss the change of dominance relationship between the front and rear legs according to high-frequency oscillation of wobbling mass through numerical simulations.

We utilize the depth visual information to present the surface detection/classification for the stable and energy-efficient locomotion of point-footed bipedal robot under the unstructured environment. The latest depth-vision systems offer very accurate 3D-point-clouds along with the RGB-images. The proposed method segments and classifies the 3D-images using machine-learning tools as random forests (RF), support-vector-machines (SVM) and relevance vector machines (RVM) for pixel-level classification and object-based image analysis (OBIA) to achieve accurate object segmentation. Contrary to the other existing methods, surface-recognition based robotic locomotion control is useful to estimate the optimal parameters to avoid unwanted-areas and to plan the collision-free energy-efficient walking. The major contribution of this work is to integrate the accurate surface modeling and locomotion rules for the bipedal walking control. Finally, we evaluate the results for average distance traveled and average energy consumption by the bipedal robot walking trajectory under different settings of unstructured environment.

This paper presents a novel transfer learning method in reinforcement learning. The method is inspired cognitive psychological theory which is self-body representation. In recent years, intelligent robot technology such as reinforcement learning has been discussed for the real world application. As a recent technique, transfer learning is proposed for reducing the learning time of reinforcement learning by reusing obtained policies. However, transfer learning is needed to be designed mapping features between body structure of robots. The mapping called inter-task mapping in transfer reinforcement learning. For example, correspondent motor and motion of the robot, and it is designed by human. In this paper, mapping method is proposed inspiring self-body representation which is suggested in cognitive psychology, further evaluate the effectiveness of proposed method in experiment with small size humanoid robot.

This paper presents decision method of policy reusing ratio of transfer learning in reinforcement learning based on gradient descent method. In recent years, learning robot system has been discussed for the actual applications. To reduce the learning time, transfer learning framework is proposed and the method is knowledge reusing mechanism. In particular, effectiveness of transfer evaluation method such as a transfer surface is proposed in reinforcement learning and adjustment method of transferring ratio is also proposed. However, decision of value of transferring ratio is depends on human intuition and experience. In this paper, automatic transferring ratio estimation method is proposed based on gradient descent method with random initial value and statistics in multiple estimation trials, further evaluation of proposed method in two different transfer surface.