Kyushu University Academic Staff Educational and Research Activities Database
List of Presentations
Satoshi Nishikawa Last modified date:2021.10.26

Associate Professor / Systems and Control / Department of Mechanical Engineering / Faculty of Engineering


Presentations
1. Mental Health Score Estimation System Based on AdaBoost.
2. QOL Estimation based on Multimodal Learning through Interaction with a Communication Agent

When a monitoring system or a communication robot for the elderly welfare interacts with a human, it is important to estimate the user's state and to generate a behavior based on it. In the field of welfare for the elderly, quality of life (QOL) is a useful indicator, not only for human physical suffering, but also for treating mental and social activities in a comprehensive manner. In this study, we propose a QOL estimation approach that integrates facial expressions, head movements, and eyes movements in the process of interaction with a communication agent. To this end, we implemented a communication agent and constructed a database based on the information collected through communication experiments with humans. In addition, we implemented a multimodal learning estimator that incorporates C3D, a three-dimensional convolution and performed learning with head fluctuation and gaze feature extraction. Our results show that multimodal learning that integrates all of facial expressions, head fluctuations, and eyes movements was realized with less error than single modal learning using each feature separately. From our experimental results, we concluded that the proposed system can be used sufficiently as a QOL estimator.

.
3. Competitive physical interaction by reinforcement learning agents using predictive models

Though there are a lot of researches about Physical Human-robot Interaction (pHRI) using prediction, few researches work on inducing the opponent's action or outwitting the opponent. We made the push-hand game environment in order to focus on generating strategic actions and tried to make reinforcement learning agents to learn these actions by adding rewards which are directly proportional to the degree of inducement (induction reward) or the degree of outwitting (outwitting reward), defined in this research. As a result, we demonstrated that the induction reward decreases the agent's predictive error and the outwitting reward increases the opponent's predictive error, and both of them didn't contribute to the winning percentage.

.
4. Throwing a Ball of the Pneumatically-Controlled Continuum Robot Arm using Reinforcement Learning with Movement Primitives

Previous research on reinforcement learning for continuum robot arms have been dealt with a relatively small number of active degrees of freedom and made experiments of simple tasks such as reaching. We aimed to learn to throw a ball by reinforcement learning in a pneumatically-controlled continuum robot arm that has nine actuators. We adopt Cost-regularized Kernel Regression (CrKR) which uses dynamic movement primitives (DMPs) which is one of the movement primitives. In the simulation, the continuum arm was able to learn how to throw a ball forward. We made the same experiment for our real continuum robot arm and found that the learning progressed in the experiment.

.
5. Reduction of the Effect of Spinning Ball in Table Tennis by High Speed Swinging using Flexible Arm

In table tennis robots, a spinning ball is difficult to be returned to the opponent court because its direction of bound is dependent on the spinning direction and amount. We aimed to solve this problem by swing motion that is robust over a large range of spinning pattern without accurate recognition of spinning. Specifically, we examined the relation between swing speed and returning direction. When the racket speed was 7 m/s, the difference of returning direction reduced to 20 % compared to no swing against the spinning ball in the range from -20 rev/s to 20 rev/s. In addition, we focused on the flexibility of swinging arm to increase the swing speed. We found that a flexible arm swung faster than a rigid arm in the whole range of motion when exploiting the joint angle limit.

.
6. Development of pneumatically actuated musculoskeletal humanoid for dolphin kick research

The dolphin kick is the swimming style with an undulation of a body conducted after start and turn. Practicing dolphin kick is one of the effective way for improving swimmers' record. It has been researched using real swimmers, simulation and robots. Dolphin kick research requires a combination of these methods. However, existing robots have inadequate points including their body size and lack of a flexible spine. Thus, we developed human-scale pneumatically actuated musculoskeletal humanoid called Triton. It has flexible spine and adjustment system for lumbar joint stiffness. We conducted preliminary experiments that compare time-series of thrust force and joint angle between Triton and experimental results with simulations and human data.

.
7. Yusuke Arai, Satoshi Nishikawa, Ryuma Niiyama, Yasuo Kuniyoshi, Compliant Jumping Mechanism with Bi-stable Structure, the International Conference on Robotics and Automation (ICRA 2017) Workshop on Advanced Fabrication and Morphological Computation for Soft Robotics, 2017.05.
8. Development of humanoid robot arm for badminton with high accelerration and high speed wrist

Badminton needs dynamic and high speed motion, so previous humanoid robots were difficult to perform like human. For entertaining or training humanoid badminton robot in the future, we propose new robot arm which consists of structure integrated pneumatic cable cylinder and noninterfering many DoFs joints. We made a real robot arm with high acceleration and high speed wrist and conducted some performance evaluation test. Pronation joint achieved 32.4rad/s in 90 deg, which surpasses human wrist speed in badminton smash. It has also good reproducibility for dynamic feed-forward control or learning control. We also make it hit actual shuttle, and shuttle's initial velocity achieved 28m/s by full swing. And it succeeded skillful shot, spin hairpin, by utilizing the humanlike wrist. This research is the first step toward humanoid badminton robot.

.
9. Control of the handspring robot with rolling-jumping locomotion

The continuous motion composed of rolling and jumping such as continuous handspring is useful as the locomotion of robots. In order to realize the continuous handspring, we propose a multi-link robot model composed of two limbs and a body and the control law for the model. We verified the controller using the combination of attack angle control and the control based on the angular momentum of the robot in the simulation. In the verification, the robot model realized 100,000 steps in the continuous handspring in the certain environment. In addition, it realized 20 steps after the temporary change of coefficient of static friction of the floor from infinity to four and over in the verification of the robustness against the temporary disturbance.

.
10. Development of a structure integrated pneumatic cable cylinder robot executing a jump-hit motion

Humanoid robots must execute situation-adaptive whole-body dynamic movements to protect humans. However, design of such robots have been unclear. Thus, our aim is to clarifying a design method of a robot executing a jump-hit motion representing these movements. For such a robot, we developed a light, high power and large stroke pneumatic actuation system equipped with a cable cylinder integrated in a link structure and made of custom-made plastic parts. We developed an arm robot, a leg robot, and an arm-equipped legged robot with this system and evaluated their physical performance. In experiments, the developed robot executed different jump-hit motions against a ball flying from the different directions.

.
11. Reservoir Computing with Body Structures and Dynamics

Some actual creatures calculate not only on their brain but also on their body itself. This is called morphological computation. Physical reservoir computing, which uses body itself as a neural network like system, is one approach of the way of morphological computation. The purpose of this research is to understand the effect of body structure and environment around it by the experiments with Tensegrity structures. When spring constant was large, structures were hard to move and dropped its ability. When input amplitude was large, input could counteract the effect of past input and dropped its ability. In addition, the environment around the body especially how to contact with ground (coefficient of restitution) gave large effect on calculation. With ground, the convergence oscillation's period and amplitude changed.

.
12. 1A1-A08 Effect of Bi-articular Spring-damper System on Disturbed Pole Vault(Mechanism and Control for Wire Actuation System)
Pole vault is an interesting task because of its high performance using a large elastic element. Although previously we showed that active bending motion is effective to improve pole vault height, it was sensitive to disturbance. In order to stabilize pole vaulting performance, physical feedback is important because of its quick response. Therefore, we focused on musculoskeletal system, especially its bi-articular spring-damper system, and investigated its contribution to stabilization of pole vaulting. We compared pole vaulting performance of a non spring-damper model, a mono-articular spring-damper model and a bi-articular spring-damper model in simulation on the disturbed condition. As a result, we found that the mono-articular spring-damper model and the bi-articular spring-damper model tended to be reduce negative effect of large disturbance at early timing of pole bending motion. In addition, bi-articular spring-damper model had intermediate response characteristics to disturbance between other two models. These results implicate that musculoskeletal system contributes to stabilize pole vault..
13. 1P1-P03 Preliminary motions for minimizing transition time to probabilistically-variable target states(New Control Theory and Motion Control (1))
Rapid reacting motions should be needed for robots working in human environments. Preliminary motions are expected to improve speed of such motions. Thus, we investigated preliminary motions in such complex environments. In this investigation, we calculated expected value of minimum time for state transition of a simple robotic arm model. In addition, we compared the value of cyclic motions with stopping as preliminary motions. As a result, we found that stopping states make speed of reacting motions more fast when target states can be predicted enough and cyclic motions may make the speed more fast when it is difficult to predict target states..
14. Body-Pole Coordinated Motion from the Pole Bending Perspective in Pole Vault.
15. Resisting motion to improve vaulting performance in pole vault.
16. Musculoskeletal Quadrupedal Robot with Variable Distribution Ratio of Joint Torque.
17. 筋骨格系駆動のヒト規範足部を備えたロボットによる跳躍.
18. Motor Learning by Phase Division with Sparse Coding of Muscle Activation.
19. Satoshi Nishikawa, Yasunori Yamada, Kazuya Shida, Yasuo Kuniyoshi, Dynamic Motions by a Quadruped Musculoskeletal Robot with Angle-Dependent Moment Arms, International workshop on bio-inspired robots, 2011.04.
20. Yasunori Yamada, Satoshi Nishikawa, Kazuya Shida, Yasuo Kuniyoshi, Emergent Locomotion Patterns from a Quadruped Pneumatic Musculoskeletal Robot with Spinobulbar Model, International workshop on bio-inspired robots, 2011.04.
21. 床反力制御による筋骨格ヒト型ロボットの走行.
22. 1A2-A07 Muscle Activation Pattern Control for Running of Musculoskeletal Robot
Despite the considerable complexity of the human musculoskeletal system, human beings are capable to move with great dexterity and ease even under instable conditions as during running. In order to understand the control mechanisms underlying the generation of such skill-full behaviors, we propose to study running, a highly dynamic motion, using an anthropomorphic bipedal robot. To do so, we propose a method called Muscle Activation Pattern Control based on physiological observations of athlete runners that we apply in a simulation of our running robot. Using this method, we show that the robot can autonomously acquire patterns of running motion over 3m/s..