It recognizes and controls the transported object of the drone equipped with the iris hand. This hand has many suitable features as a drone end effector. Recognition accuracy and coordinates are important when recognizing and grasping a transported object. Therefore, in this research, we introduce a position estimation system using a camera and a communication system using ROS, and construct and control a system assuming automatic transportation in an indoor environment.

Multicopters are useful in a variety of fields. However, there is still a risk of loss of control due to updrafts or wind generated by the multicopter itself. Therefore, the goal of this research is to achieve robust flight against wind disturbances. As a first step, we develop an onboard airflow visualization system to visualize the airflow around a propeller in flight in order to understand aerodynamic phenomena. By conducting ground experiments using a wind generator, we show that it is possible to visualize the airflow generated around a propeller. Furthermore, it is shown that it is possible to evaluate the flow field through PIV analysis of the images taken by the onboard camera.

Our team developed mechanism that aims securing regidity and reducing weight of UAV by constructing with box-shape objects. It is a mechanism that feeds out box-shaped objects one by one, and the transfer efficiency is improved by performing multiple transfers. In this paper, we describe the trial production and operation test of the individual feeding mechanism, and confirm that the rigidity is secured.

Contact force control of a multi-rotor helicopter allows for achieving aerial tasks that require contact motions with objects．A quadrotor helicopter control system is an underactuated system．It is necessary to tilt the attitude angle of the quad rotor，when the quad rotor exerts thrust in the horizontal direction with respect to the ground. Therefore, there is a problem that the response of the quadrotor contact force control system is slow．In this paper，a contact force control system for a fully-actuated 8 degrees-of-freedom quad tilt-rotor is proposed，and its validity is verified by simulations．

Recently, UAV are researched and used in various fields. Especially, UAV autopilot to reduce labor costs and fill the gap in skills is an essential technology for industrial applications. Autopilot, especially terrain following requires not only longitude and latitude but also altitude information on the map. GCS software, such as Mission Planner and MAVProxy support terrain following. There are various methods for obtaining altitude information, such as using sensors such as GPS and Lidar, and one of them is DEM that expresses the topography of the ground surface using digital data. In this study, using DEM, I performed highly accurate terrain following flight.

In this research and development, we will develop the performance evaluation criteria of the aircraft and the evaluation method to realize the establishment of industrial standards for formulating the safety standards required for unmanned aerial vehicles, and also test the evaluation process and develop the developed performance. We report on the technical verification of the evaluation technology and the validity of the evaluation method.

In transporting a heavy load, instead of utilizing a conventional unmanned aerial vehicle (UAV), this research introduces the new scalable UAV design for lifting and transporting applications. The main concept focuses on reconfigurable control layers with multiple UAVs arranged in hierarchical order. Data transmission across layers is responsible for sending desired actuator outputs to the mixer located in the low-layer UAV units. Moreover, in this scalable UAV platform, each layer has different control methods for each unique controlling task. With several small propeller configurations, it gains flight safety, stability and mobility while achieving high thrust for heavy load lifting application.

With recent developments in UAV technology, manned work has been replaced by multi-copters. Furthermore, the integration of cameras and sensors has made it possible to fly beyond visual line-of-sight, and UAVs are expected to be used for transporting. However, transporting with a single UAV becomes larger and more expensive due to the payload problem. Therefore, we propose a cooperative transportation system using multiple Coaxial-UAVs. In this paper, we control the attitude of the Coaxial-UAVs using STA architecture. The STA architecture uses Neural-Network to learn motions according to the state of the environment without considering the overall system modeling and the positional relationship between the UAVs. In the experiment, we investigate the attitude control using multiple Coaxial-UAVs with the test bench limited to 1 DoF. As a result, it was suggested that the attitude control is possible even in the asymmetrical balance of the UAVs.

This paper designs a flight maneuver that minimizes the battery power consumption during take-off or landing for a Quad-plane UAV (Unmanned Aerial Vehicle), through numerical optimization. The power consumption is calculated from the inverse dynamics and the rotor power model for transition flight between a rotary-wing mode and a fixed-wing mode. In a maneuver designed for landing, the UAV is decelerated due to drag force of a wind-milling propeller.

UAVs are used to conduct status surveys in places where people cannot reach in the event of a disaster, but it is difficult to fly a ordinary UAV in bad weather such as strong rain or strong winds. In this research, we developed a quad copter with tilting mechanism that can fly even in bad weather. We will report the relationship between the attitude the experimental aircraft due to the tilt angle of rotors.

A cooperative tracking system using the LiDARs on the ground and cameras mounted on UAVs is proposed. With the conventional system, it was not possible to instruct the shooting a person existing in the occluded area of LiDAR. In this paper, a sensor node embedding LiDAR, RTK-GPS and IMU sensors is developed, and a method for complementing the shooting position when the occlusion occurs is implemented.

Multicopter is capable of vertical takeoff and landing, and it does not require a long runway. Therefore, in the event of an emergency such as a disaster, it is possible to quickly respond to grasping the current situation by aerial photography. However, ordinary multicopters are designed for takeoff and landing in a flat surface location. So it is difficult to take off and land on rough terrain such as a place where rubble spreads or a mountain slope, and there is a possibility of falling. In order to solve this problem, a mechanism that can land at any landing point is required. Therefore, we propose a landing gear using a linear actuator that can land on rough terrain such as steps and slopes.

We have developed an inverted duct fan air vehicles. These are propulsion-type aircraft equipped with two duct fans at the rear of the aircraft and capable of hovering, flying, and VTOL while maintaining the inverted state by thrust vectoring. By introducing this hovering / flight control into a fixed-wing aircraft with a thrust vectoring mechanism, we aim to realize VTOL and hovering in an inverted state in the same way as a tail sitter.

In order to manipulate objects using aerial robots, conventional way using quadcopter with a robotic arm has two problems; Underactuation and Weak manipulation force. To solve underactuation, 1-DoF thrust vectoring mechanism is attached to each propeller and enables fully-actuated flight control. Flight control uses online non-linear optimization which takes thrust forces and vectoring angles as inputs. To increase manipulation force, this paper proposes whole-body manipulation by multilinked robot taking advantage of environmental adherence. Motion planning for manipulation which includes contact-detection, mode transition, end-effector position control was proposed. The robot successfully opened the door in real machine experiment, and showed its ability to manipulate large objects.

A new voice interactive system to control an assistant manipulator for laparoscopic surgery performed by a surgeon in a sterilized area was proposed. The voice control system consists of a microphone, a limited word dictionary applied to speech recognition engine Julius, and a forceps manipulator LODEM as the third arm. We propose the voice interactive system with a rough drive mode of clock position and a precise drive mode of moving a bit after confirmation from the surgeon. The recognition rate of the dictionary and a motor drive using the rough mode was confirmed.

Echography is an imaging diagnostic method that has advantages of low cost, minimally invasive, real time, compactness, and simplicity. It is indispensable equipment in every medical field but examination with echography greatly depends on the expertise and experience of an operator. In this study, we propose a new robotic-echography support system (Echography Assisted Robotics, EARs) using a robotic platform ROS (Robot Operating System).In this paper, we constructed an echography environment with ROS and reproduced the movement of the ultrasound probe with a robot. We also conducted experiments to measure and reproduce the probe scanning trajectory, and compared the acquired echogram. As a result of our experiments, we succeeded in acquiring echogram by using EARs, but there were cases where the EARs could not handle the motion of the patient’s body surface, resulting in different diagnostic images.

In recent years, the number of people who need medical care has been increasing in an aging society with a shrinking population. However, in spite of the decreasing number of medical personnel, it is necessary to provide high-level medical surgery in a safe and secure manner. Therefore, research and development of surgical robots and manipulators are being promoted to improve medical technology, reduce medical accidents, and reduce the burden on surgeons. In this study, we propose a new roboticechography support system (Echography Assisted Robotics, EARs) using a robotic platform ROS (Robot Operating System). In this paper, we conducted experiments to orientation of the probe, and conducted an accuracy comparison experiment with a motion capture system EARs with control parameters suitable for cooperative motion and using ultraleap. As a result of our experiments, we confirmed that the positional accuracy of the end-effector is higher than that by motion capture.

Since laparoscopic surgery is performed through the images displayed on the monitor, it requires a higher degree of surgeons’ skills than open surgery, in which surgery is performed while directly checking the surgical field. In order to support decisions and surgical procedures in laparoscopic surgery, a surgical support system that displays deformation of organs caused by contact with a surgical instrument calculated by a finite element method has been developed. The usefulness was proved by the operation verification with a simple elastic body and the deformation and the average error within 1 mm. After that, the state of contact and deformation of the virtual organ model was confirmed.

This paper presents a concept of Avatar Foot aimed at remote palpation and treatment of the ankle joint, which is a device intended to reproduce the motion characteristics of a patient's ankle joint. As the first step to realize this device, we propose a set of pneumatic flexible actuators suitable for reproducing the muscle characteristics, and verify the validity of the proposed method by experiments.

In recent years, a surgical method called IVR (Interventional Radiology) has attracted attention. It is applied to cancer treatment and biopsy, and has the feature of less burden on patients. However, radiation exposure of doctors has become a problem. Therefore, I has been developed the puncture robot that can be operated remotely. The puncture robot has made it possible to reduce the radiation exposure of doctors to zero. Currently, automatic puncture is a major goal of robots. One of the requirements for the realization of automatic puncture is the detection of the posture information of the puncture needle from the CT images. This paper proposes a method for detecting the tip coordinates and orientation of the needle from CT images.

The operation of medical instruments, such as endoscopy, relies heavily on the physician's senses, and simulators and remote control should be accompanied by the presentation of force sensation. For accurate force presentation, it is necessary to convert the grasping sensation into data with vector force sensing which does not disturb operation. In this study, we have adopted our ultra-small tactile sensor which can detect both normal and shear forces to measure force during grasping and twisting of the tube model.

Under laparoscopic surgery, surgeons insert forceps from small incisions and operate. Therefore, the motions of forceps are restricted. To design surgical robots, we developed the simulator for surgical robots with linear motion unit via trocar. In this simulator, the linear motion unit shortens depending on the arm angle and manipulability index. To evaluate the automatic adjusting linear motion unit, we developed another simulator, without the linear motion, to compare. From the comparison, we figured out that higher manipulability was obtained by the automatic adjusting. Furthermore, we evaluated the position of trocars. We compared the trocar positions with manipulability index and figured out that 0 or 10 degree position would be better to do the work under this simulation conditions.

Gastrointestinal cancer can be treated by accessing the lesion through the natural orifice by a flexible endoscope if it’s detected in its early stage. However, the endoscopic treatment requires an advanced technique because of the limitation of instruments in dexterity. In this study, we have developed a novel multi-DOF endoluminal forceps that is sized to be inserted through the working channel of conventional flexible endoscope. The proposed forceps has 2-DOF of bending and 1-DOF of grasping by wire-driven mechanism. The proposed structure is a monolithic structure with a compliant mechanism, which is advantageous for mass production and assembly. The current prototype made by a 3D printer, showed a positive feasibility in its preliminary motion test. Finite element analysis (FEA) revealed that the proposed mechanism can be bent 90° without breaking.

In this paper, we used a flexible wearable tactile sensor that utilizes acoustic reflection, and investigated hardness estimation based on the sensor output using human spontaneous haptic exploratory motion. Machine learning methods with support vector machine and convolutional neural networks were adopted to the acquired sensor output for the identification of hardness of cervical models. As a result, the correct answer rates of both learning models were 100 ％ for the soft model and about 80 ％ for the hard and medium models.

In this paper, we propose a novel hydrogel swab for specimen sampling, which has a swab tip made of superabsorbent hydrogel and a detachable fiberscope to enable observation of the nasal cavity for a reliable sampling of a large amount of nasopharyngeal swab fluid. A hollow probe with a 0.64 mm diameter fiberscope inside was fabricated using a 3D printer. The bending stiffness of the probe was designed to be comparable to that of the commercial swab. The tip of the swab was fabricated using superabsorbent sodium polyacrylate and photo-crosslinkable water-soluble polyurethane. In a 10-second water-absorption experiment using a methylene blue solution, this swab provided about 2.5 times as much water as a commercial swab. The concentration of methylene blue solution released into the virus transport solution was 5.7 times higher than that of the commercial swab. From these results, we confirmed the effectiveness of the prepared swabs.

It may be necessary to estimate the muscle exertion torque of the ankle joint in lower limb movement support using a motor and automatic evaluation of the degree of spasticity using a motor. Although method for estimating ankle joint torque using surface electromyogram and ankle joint angle have been proposed so far, there is a problem that an estimation error occurs when the knee joint angle changes. Therefore, in this paper, we propose an ankle joint exertion torque estimation method using surface electromyogram and ankle knee joint angle. The effectiveness of the proposed method is verified by experiments.

In order to realise force sensing in robotic minimally invasive surgery, we developed a forceps mechanism with an optical fiber sensor integrated into the structure. First, we designed the forceps that can incorporate with the sensor, then conducted a temperature characteristic evaluation and a loading experiment to verify that the designed component can properly measure the load. These preliminary evaluations showed promising results to be further implemented in a surgical robot.

Using nasal endoscopic surgery as a case study, we are investigating the pose estimation of surgical tools using endoscopic images. Current state-of-the-art algorithms are well known to require a large amount of training data. Our group has been investigating the generation of synthetic images with realistic appearance to consistently address this difficulty using photorealistic rendering. In this work, we preliminarily investigate the use of generative-adversarial networks to further improve the rendered images' appearance. Our current implementation reveals some difficulties in this strategy and highlights directions for future work.

Although the utilization of robotic forceps has reduced technical burden on surgeons from its constrained and difficult procedures, challenges for currently used robotic forceps such as their complicated structure, low durability, and high production and maintenance costs still remain to be solved. Regarding these challenges, in this study, we propose a conceptual design of a component of a compliant disposable robotic forceps with multi-degree-of-freedom for minimally invasive surgery. The robotic forceps are made of polymer materials, and the parts performing grasping movement consist of symmetrically arranged springs. This paper describes the mechanism of the forceps and evaluations using Finite Element Analysis. The evaluations reveal that the design with double springs gives lower stress distribution on the structure than ones with single springs while exerting same grasping force, which is key function for forceps.

In this study, we realized in-air knotting of rope with the dual-arm two-fingers robot by using deep predictive learning. By training two deep neural networks (CAE and LSTM) with robots’ sensorimotor experiences including visual and proximity sensors, we had the robot perform bowknot and overhand knot while dynamically responding to the state of the rope online. In addition, since we designed task motion based on the Ian knot method by using the dual-arm two-fingers robot, our method does not require a dedicated workbench or robot hand. As a result of the experiments, it was confirmed that the robot could perform tasks with high accuracy even for the unlearned rope.

In this study, we propose a robotic buttoning task by a dual-arm humanoid robot using marker-based recognition. In our method, we design buttoning task by dividing into multiple subtasks, and the robot performs them in order. When the robot performed each subtask, it conducts and adjusts motions based on the position of the button and its hole obtained by maker-based recognition. For the implementation of the robotic buttoning task, we used a Baxter, an experimental T-shirt, and a button-aid. As a result of experiments, although the problem of low success rate remains, the robot was able to realize a series of buttoning operations.

We propose robotic buttoning task by a dual-arm robot driven by deep neural networks (DNNs). The previous study indicated the robot was able to perform buttoning tasks by marker-based recognition; however, some issues remain. In this study, we trained two DNNs; CAE and LSTM with the robot’s task experiences consist of camera images and joint angles of the robot. We verified our method through the simulator and real robot experiments. The method generated motion online based on sensor information acquired from the robot, and the robot successfully performed the buttoning task. In addition, the results showed our method could generate smooth motion based on only one viewpoint.

In object manipulation involving contact, the object may be hidden from the camera, and haptics are useful to compensate for this information. In this study, we propose a deep learning-based method for generating robot motions using tactile data. We introduced attention mechanism for image feature extraction, softmax transformation for motion generation, and convolutional neural network for processing tactile sensor data. We tested the effectiveness of the proposed method on the unzip task of an flexible bag. We confirmed that the proposed method can realize the motion generation according to the deformation of the zipper while reducing the load on the zipper, and achieved a success rate of 90 percent.

In order to realize cooperative work between humans and robots, we developed a technology that understands human instructions and generates appropriate motions based on learning experience. In this paper, we proposed a motion generation method that uses an attention mechanism to extract object location information from visual information and an association mechanism to predict the indicated object. In a situation where multiple objects are placed, we confirmed that the robot accurately grasps the object indicated by a person.

In the service field, it is important not only for efficiency but also for the impression that the interaction with the robot gives to people. Therefore, the purpose of this study is to generate meaningful motions for communication robots. In this study, we extract the features of motions that express a certain meaning from classical dance motions that have different characteristics depending on the region. We are trying to solve the difficulty of data analysis, including small number of data and many dimensions, using sparse modeling. By extracting features using sparse modeling and classifying the dance motion data using the features, we examined the validity of the extracted features. First, we tried to extract the features of the motions that express the meaning, but we have not yet succeeded in extracting them. On the other hand, it was found that it is possible to extract common features between dances with regional characteristics.

We are aiming to develop a robot that plays “play tag.” The purpose of this study is to propose a method for a robot to approach a person with the intention of touching the person. Specifically, we set up a circle around the person and propose a method that leaves along the tangent line of the circle and a method that selects the contact point according to the movement of the person to be followed.

Dilated pupils make favorable impression. Focusing on the benefits of the dilated pupils, we have developed a pupil response pet-robot for improving familiarity. This robot can enhance familiarity by enlarging pupils in synchronization with the body contact. In this study, focusing on creating an atmosphere that makes it easy to talk with the pet-robot for promoting self-disclosure during communication, we investigated the size of the pupils that give favorable impression with or without speech input. Based on the experimental results, we developed a speech-driven pupil response pet-robot that generates both a clearly enlarged pupil response and a blink that is synchronized with the speech rhythm.

In the field of motor control, ”Synergy Hypothesis” is a concept that behaviors are constructed by a combination of the patterns of movement called synergy. A large number of studies have supported this hypothesis as an explanation for redundant body control in humans and animals. Although much research has been done on the concept of synergy, the mechanism of synergy generation remains still unknown. In this study, we generate gait by Model-Based Reinforcement Learning, which is a control framework utilizing a function like internal model in the human brain, and analyze the spatiotemporal synergy inherent in the generated gait. As a result, we found that walking speed and energy efficiency are important factors for synergy generation.

A patient who temporarily uses a crutch does not usually know the proper using method of a crutch, and an improper usage may lead to the secondary accident like a falling. In this research, a crutch walk training system where the trainee wears a HMD was proposed. The trainee was required to stop walking stably when a virtual obstacle is presented on a HMD. In the first experiments, it was found that the acceleration of the participant’s trunk became irregular and it took more time until the participant stopped walking stably after an obstacle was presented at the early instant of the swing phase of the healthy limb. In the second experiments, 6 of the 9 participants acquired a proper walking method to stop walking stably. These results showed that the adaptability of the participant was improved through trainings with the developed training system.

In this paper, we propose a novel graph-convolution-based network architecture for 2D-to-3D human pose estimation using multi-scale and multi-level features. For multi-scale features, we propose a novel pooling/unpooling approach considering the human skeletal structure to obtain both global and local feature information of human pose, and then use it to compose an encoder-decoder structure applicable to the graph-structured data. For multi-level features, we utilize the feature information in each depth-level in the neural network architecture, thus enhancing the performance of the model. Extensive experiments show that our proposed model architecture using multi-scale multi-level features outperforms the state-of-the-art approaches.

Japanese “Bunraku” puppets have attracted people by showing their impressive and emotional action, even though the structures and motions are different from human beings. The exaggerating and omission of motions, which include expansion and contraction, and high frequency shakings of the body, are effectively utilized to express humanistic emotions. This paper describes a novel humanoid robot which has mechanisms to reproduce functions and motions of Bunraku puppets by extracting the mechanism of a Bunraku puppet and operators. Motions of real Bunraku puppet were captured optically and processed by using kinematics to calculate the reference angles of every joint of the robot. Both motion data that include and exclude the exaggerated motions such as expansion and contraction of neck and chest were prepared and executed using the developed humanoid robot to be compared. It has been clearly shown that the motion with exaggeration is effective to express humanistic emotions.

Due to growing needs for labor-saving, a large number of logistics support robots for transportation are expected to coexist and collaborate with workers in future logistics sites. Those robots should be introduced while avoiding the decrease of transfer throughput due to interference between humans and robots or among robots. In this paper, we constructed a simulator that allows measuring the performance of multiple logistics support robots working in a collaborative environment with human workers. The simulation results showed that transport efficiency actually decreases as the number of robots increases. We then identified several factors causing the reduced performance by analysis of the results to propose countermeasures for each factor.

When performing multiple autonomous mobile robots simultaneously, if the planned schedule gets confused, collision and deadlock will occur. Operation control system mount path planning unit for generating efficient traffic route. Passage permission was given to the autonomous mobile robot in crossing this side, and the continuous duty with high robustness was realized because traffic management system controls an intersectional passage order.

When horses walk, they move their necks as well as legs rhythmically. In addition, horses exhibit versatile neck movements according to locomotion speed. This phenomenon could be a key to understanding how animals orchestrate their large degrees of freedom in their bodies. However, the mechanism underlying this phenomenon still remains elusive. In the present study, we propose a decentralized control mechanism for the coordination of neck and limb movements in which inclination of the body trunk is fed back into neck and limb motions. Using a simple two-dimensional physical simulator we developed, we successfully reproduced speed-dependent neck movements during walk and trot gaits.

The goal of this research is to realize autonomous distributed cooperative behavior among multiple robots in the same area. To achieve this, it is necessary to establish (1) a stable autonomous decentralized communication network that can dynamically respond to increases and decreases in the number of robots, and (2) an autonomous decentralized control method that allows each robot to autonomously estimate other robots’ positions and perform cooperative behavior. In this paper, we study a method for estimating the relative positions of robots based on the RSSI values obtained from their communications. We also study the formation control using this relative position estimation method.

Autonomous decentralized control could be the key to design soft-bodied robots that have a huge number of degrees of freedom. To address this issue, we focus on a wandering spider with a hydrostatic skeleton and aim to understand the decentralized control mechanism underlying their coordination pattern between limbs (i.e., interlimb coordination). For this purpose, we employed an approach wherein we construct a simple mathematical model based on biological insights. We proposed a phenomenological model that can describe fluid dynamics and built a three-dimensional robot model with a simple local reflexive mechanism based on the interaction between flexor muscles and body fluid in their legs. As a first step, we succeeded in reproducing locomotion in which two legs move in anti-phase via simulation.

Understanding control mechanisms underlying adaptive animal locomotion will provide us a design scheme for robots which can behave autonomously depending on environments. To address this issue, it is important to capture the interplay between local pattern generating circuits, descending commands from higher centers, and sensory feedback. However, the essential interplay between them that generates adaptive behavior in response to the situation remains unclear. Here, we focused on centipedes and explored the interplay by observing the decision-making when they faced the edge of the precipice during walking. We found that a ventral nerve cord-transected centipede stopped walking in a shorter distance forward than the intact one. From the results, we extended our previous mathematical model for walking based on decentralized control and proposed a descending control mechanism which utilizes ground reaction forces detected at the legs for deciding whether to move forward.

Myriapods such as centipedes and millipedes can coordinate their elongate multi-legged body to locomote adaptively in unstructured and unpredictably changing environments. Clarifying the underlying motor control mechanism for locomotion will help to realize multi-legged robots that can move adaptively in various environments. However, the interlimb coordination mechanism that can generate load-dependent adaptive gait inherent in myriapods has not yet been elucidated. To solve this problem, we focused on the dynamic muscle properties as actuators for walking which may be closely related to the generation of gait patterns. In this paper, we observed the response to the load on the body and found that centipedes change their gait and increase their duty cycle. To understand this behavioral finding, we proposed a mathematical model for the locomotion control in which the force-velocity relationship of leg muscles was taken into account.

Quadrupeds exhibit versatile gait patterns according to locomotion speed and environment. In particular, they exhibit jumping motion during high-speed running. To investigate how the gait patterns and jumping motion are generated, various intraspinal neural network model have been proposed. However, high-speed running on uneven terrain was not realized. To address the problem, we focused on the vertical velocity of the body parts and designed a new leg-control model that realizes jump by considering leg control during the swing phase. As simulation results, we succeeded in widening the range of locomotion speed and running on uneven terrain. In this study, we developed a one-legged hopping robot and verified the proposed model in the real world.

There are increasing demands for developing systems consisting of multiple mobile agents that can move quickly, smoothly, and safely. Although several models in which agents move around on a two-dimentional field while avoiding others have been proposed, the proposed control schemes could not satisfy the above requirements. To address this issue, our group previously developed a mathematical model in which prediction of future motions of nearby agents is added to the social force model. However, this model requires a considerable calculation cost. To tackle this problem, we improve our previous model by incorporating active sensing. In this model, each agent has an internal state called “collision risk level”, and its sensing direction is determined on this basis. It is expected that the proposed control mechanism enables agents to reduce the calculation cost while achieving the above-mentioned three requirements.