If you have any difficulty in finding or downloading papers listed below, please feel free to contact me via Email or Google Form.
The following contents may be updated.
Frozen shoulder, also known as “adhesive capsulitis” or "periarthritis of the shoulder", is a syndrome that causes shoulder pain and limitation of the upper extremity range of motion (ROM), and it is estimated that 2-5% of the total population has experienced it for at least once.
The treatment of frozen shoulder can be mainly divided into 3 periods: (1) freezing stage lasts for 2-9 months; (2) frozen stage that lasts for 4–12 months; and (3) thawing stage that lasts for 5–26 months, and it is important to perform physical therapy during the frozen stage. In comparison with conventional types of physical therapy, Continuous Passive Motion (CPM) therapy has shown advantages in pain reduction. In this therapy, the physical therapist holds the patient's upper arm and performs repetitive movements of the upper arm and shoulder within the patient's joint ROM.
However, it is not difficult to imagine the heavy workload and fatigue for physical therapists during the months of repeating CPM therapy. Therefore, CPM machines have emerged as partial substitutes for physical therapists in terms of performing CPM since 1970's, contributing to faster recovery of patients.
Nevertheless, the CPM machines developed and sold on the market so far cannot completely replace the role of physical therapists and perform customized CPM for a patient according to their individual features. Specifically, the following two issues still require further improvements or even solutions: (a) no independent motion training is available for the scapula (the bone of the shoulder blade); and (b) the CPM machines currently on the market cannot exactly reproduce the arbitrary training motion performed by a physical therapist.
Consequently, a rehabilitation robot for frozen shoulder that improves and solves the issues (a) and (b) described above was developed (the figure above). The training motions of both shoulder joint and scapula are enabled by the robot link designs and reproducibility of an arbitrary training motion by a physical therapist is guaranteed by its teaching and playback mode, which is inspired by “teach and playback” of industrial robots. For the first time, the patient will be equipped with the robot and the robot records the training motions performed by the physical therapist in teaching mode. Once the motions are recorded, the robot can reproduce the exact same motions in playback mode without the help of a physical therapist. Meanwhile, the robot can be applied to both the left and right shoulder of a patient by rotating 180[deg].
Specifically, as the figure below shows, there are 6 Degree of Freedom (DoF) for both shoulder joint and scapula in total, and it was validated that the developed rehabilitation robot can accurately (instantaneous maximum joint angle error of 5[deg]) reproduce all 12 DoFs for both the left and right shoulder of a healthy subject. Moreover, to simulate an arbitrary training motion by a physical therapist in actual situations, a motion pattern involving multiple DoFs was also performed for both the left and right shoulder and similar results were obtained.
*Screenshots of experiments can be seen in the Appendix of the paper shown below.
Click here for paper link
[1] Sun, X., Makino, K., Kurita, D., Kaneko, H., Ishida, K., & Terada, H. (2025). Development of a Rehabilitation Apparatus for Frozen Shoulder Enabling Total Motion of Shoulder Complex. Robotics, 14(2), 12.
"Quad-SCARA" (figure at upper left) platform was developed by Terada Lab., University of Yamanashi, Japan and it aims for grasping and manipulation of soft materials like a handkerchief. It consists of four SCARA (Selective Compliance Assembly Robot Arm, figure at the lower left) that was developed by the University of Yamanashi and has been widely applied all over the world.
However, motion planning of these four robots in such narrow space without collision is never an easy work. Therefore, a collision avoidance approach called "Buffered Voronoi Cell" (BVC in short) is applied to Quad-SCARA, which was proposed in 2017 and has been highly evaluated and applied in recent years. Unfortunately, original BVC is targeted at multi-agent systems where each agent can move freely, and is not intended for robot arms fixed on a table. Therefore, the new concept of "static BVC" is proposed and utilized to avoid self-collision of each robot in Quad-SCARA, and parameters of BVC are optimized to enable collision avoidance to the fullest in limited space [1]. With modified and optimized BVC approach described above, folding and spreading of a handkerchief have been realized by Quad-SCARA.
[2] Sun, X., Ishida, K., Makino, K., Shibayama, K., & Terada, H. (2023). Development of the “Quad-SCARA” platform and its collision avoidance based on Buffered Voronoi Cell. Robotica, 41(12), 3687-3701. doi:10.1017/S0263574723001236
As a simulation of remote manipulation under extreme environments like disaster sites, human-sized four-limbed robot WAREC-1R is used to grasp a electric drill. Its remote manipulation (figure at the upper left) is enabled by remote control device FST, FSG (Flexible Sensor Tube, Flexible Sensor Glove, provided by Namiki Lab., Chiba University, Japan) and an HMD (Head Mount Display).
However, success rate of remote grasping of the drill is very low due to complicated shape of the drill, delay of the remote system and the difference of depth sense between reality and visuals displayed in HMD. As its solution, modelling of the robot, the camera equipped at robot hand, the electric drill to grasp and backgrounds is created in simulator Gazebo in Linux OS. DQN (Deep Q Network), an approach of deep reinforcement learning is utilized for the learning of appropriate position and orientation of robot hand to grasp the target drill based on camera view of robot hand (figure at the lower left). To reduce the difference between reality and simulation (figure at the bottom) as much as possible, the backgrounds in camera view as well as light and shadow are randomly added in simulator during learning to improve robustness of the learned model.
Moreover, object detection algorithm YOLOv3 is applied to detect the target drill in reality, so that an assist system can be developed that only requires a start command from operator as input and is capable of performing all the remaining automatically, including recognizing, approaching and grasping of the target drill. With this system, the success rate of grasping rises up from 4% to 76%.
Click here for paper link
[3] Sun, X., Naito, H., Namiki, A., Liu, Y., Matsuzawa, T., & Takanishi, A. (2022). Assist system for remote manipulation of electric drills by the robot “WAREC-1R” using deep reinforcement learning. Robotica, 40(2), 365-376.
In ladder climbing by WAREC-1 shown in the column below, end-effectors may fail to hook ladder rungs correctly due to the deformation of robot caused by its size (1.7m height) and weight (more than 150kg), and climbing fails as well. In my Ph.D. thesis, this error caused by deformation can be absorbed by sensing, yet it also takes considerable time, which slows down the speed of ladder climbing.
Therefore, as a fundamental solution, locations where deformation of the robot concentrates on are revealed by FEA (Finite Element Analysis) and motion capture technique. With pinpoint design change and addition of corresponding mechanical parts, the deformation at robot end-effectors decreased by 60.4% at the expense of 1.7% increase in total weight. After this stiffness improvement, a new ladder climbing gait called "one rung skipping 3-point contact ladder climbing" becomes available. Former necessary steps that take extra time are no loger needed, including sensing of ladder rungs for deformation compensation and stabilization with the horizontal move of CoM (Center of Mass), thus realizing both stable and faster ladder climbing.
Click here for paper link
[4] Sun, X., Ito, A., Matsuzawa, T. et al. Limb Stiffness Improvement of the Robot WAREC-1R for a Faster and Stable New Ladder Climbing Gait. J Bionic Eng 20, 57–68 (2023).
As the summary of my study in Ph.D. course,
Proposition of stability conditions and whole-body motion generation based on it;
Stabilization of ladder climbing by reaction force control at robot end-effectors;
Interval and inclination sensing of ladder rungs by ultra small proximity sensors equipped at end-effectors and ladder climbing according to the sensing data
of the robot WAREC-1 are all integrated, with the results all validate and evaluated.
[5] Sun, X., Hashimoto, K., Hayashi, S. et al. Stable Vertical Ladder Climbing with Rung Recognition for a Four-limbed Robot. J Bionic Eng 18, 786–798 (2021).
The research in [1] regarding the rehabilitation robot for frozen shoulder was supported by JSPS KAKENHI Grant Number 22K12934.
The research in [2] regarding Quad-SCARA was partially supported by Satoshi Omura Human Resource Development Fund Project.
The robot arms used in this research are donated by YGK Corporation, Japan.
The research in [3]~[5] regarding WAREC-1 and WAREC-1R was conducted by Research Institute for Science and Engineering, Humanoid Research Institute and Future Robotics Organization of Waseda University. It is commissioned by the project "Tough Robotics Challenge" of Impulsing Paradigm Change through Disruptive Technology Program (ImPACT) that is initiated by Council for Science, Technology and Innovation via Japan Science and Technology Agency (JST).
3D CAD in this research is provided by SolidWorks Japan K. K., cables and connectors are provided by DYDEN CORPORATION and the crane used in experiments is provided by KITO CORPORATION. We would like to express our thanks to all of them.