It is mostly challenging to implement reinforcement learning due to vast search space. To address this issue, Zhang et al. proposed Value Disagreement Sampling (VDS), which sets pseudo-goals based on degree of disagreement within the multiple value functions. However, the VDS approach may not set contributory pseudo-goals to learn task objectives. In this paper, we aim to enhance the learning efficiency by sampling pseudo-goals based on the state-action distribution sampled from current policy. Simulation results demonstrate the effectiveness of the proposed approach in improving learning efficiency, especially during the later stages of the learning process.

Multi-Agent Reinforcement Learning (MARL) is a framework that utilizes reinforcement learning to simultaneously learn policies for multiple agents, such as robots, within the same environment. One concern with reinforcement learning is that stochastic behavior during learning can lead to risk for the agent. In the context of MARL, appropriately avoiding risks such as collisions between agents, is necessary. This study aims to achieve mutual risk avoidance by evaluating the overall risk of multi-agent system (MAS) when each agent has its primitive risk. The focus is on the differences in the nature of rewards and penalties in MAS. The proposed method is designed based on risk evaluation using maximum punishment. Simulation results demonstrate that the proposed method achieves more advanced risk avoidance compared to risk evaluation based on mean punishment for all agents.

In chemical plume tracing, we created a high degree-of-freedom(DOF) sensing device with 11 gas sensors to acquire odor information, and made it possible to acquire 11 dimensional data at the same time. By using the 11-dimensional data, we thought we could obtain more information about the environment than ever before. we conducted measurements and discussed the results.

Myriapods exhibit diverse locomotor patterns depending on species and situation (e.g., speed, environmental media). However, the mechanism underlying these phenomena still remains largely elusive. To tackle this problem, we postulated that an inter-species control principle underlies various types of myriapod locomotion. We have recently proposed a mathematical model that could well reproduce two distinct body–limb coordination patterns observed in centipede locomotion, which include the gait with body undulation in the yaw direction. In this study, to test the universality of the proposed body–limb coordination control principle of centipede locomotion, we designed a mathematical model of millipede locomotion, which employs body undulation in the pitch direction. In addition, we developed a multilegged robot to verify the proposed model in a real environment.

Cheetah is the fastest quadruped in the animal kingdom. To understand the mechanisms underlying the high-speed running ability, we focus on the two key characteristics of cheetah running: trunk bending and footfall pattern, known as rotary gallop. While trunk bending contributes to stride extension, the extent to which the footfall pattern contributes to high-speed running remains unclear. To address this issue, we hypothesized that the rotary gallop allows for effective utilization of trunk bending. A previous simulation study suggested that flexible trunk bending in response to the footfall pattern improves running speed. In this paper, we present a simple wheeled robot to verify our hypothesis in the real world.

This paper presents a safe adaptive coverage controller for multi-agent systems with actuator faults. The centroidal Voronoi tessellation (CVT) is applied to generate an optimal configuration of multi-agent systems for covering an area of interest. As conventional CVT based controller cannot prevent collisions between agents with non-zero size, an adaptive control barrier function (aCBF) based controller is developed to ensure collision avoidance and to deal with system uncertainties. The proposed controller is verified under simulations.

In recent years, as computer technology and other technologies have become more sophisticated and faster, information has become more diverse and voluminous, and centralized management of information has become more common. However, while centralized information management systems such as those currently in use enable centralized management of information, they lack diversity, flexibility, and fault tolerance, and this has become a major detriment to more complex, large-scale systems. Therefore, it is considered most efficient to introduce an autonomous decentralized system that has high reliability and processing capacity, can change quickly and flexibly, and is used in combination with a centralized management system. In this study, we focused on autonomous decentralized systems and evaluated the mechanical and electronic control sections of a prototype four-wheeled vehicle in order to conduct an actual machine-based study.

This paper proposes a quantitative performance metric for patrolling missions by robotic swarms. Real-world robotics applications require predictability from human operators at a base station to prevent unexpected robotic behaviors and resultant harmful outputs. However, functions to ensure predictability, such as frequent communication to the base station, may reduce mission performance. The proposed performance metric quantifies the degree of situational awareness at the base station, which is strongly related to predictability. The metric visualizes the advantage of the functions for predictability enhancement and enables trade-off studies between mission performance and predictability from humans. Simulated patrolling missions with two different patrol algorithms exemplified the usage of the proposed metric.

The purpose of this study is to design a dynamic collaborative sediment transport system with backhoes and dump trucks in an unlimited environment. In an unlimited environment, it is necessary to evaluate team performance in real-time and flexibly reorganize the team structure, since it is expected that anomalies such as stuck or broken construction machines will occur and the work will not proceed as scheduled. In this study, we propose a system that dynamically and autonomously reorganizes teams by evaluating the efficiency of sediment transport per unit of time as performance. In this paper, we design a dynamic team organization flow based on the negotiation between backhoes as a team organization in a multi-team sediment transport, and examine its validity through simulations.

Cooperative transport systems by swarm robots are expected to handle a wide variety of objects. However, no studies have been reported for cases where a robot is stuck or the load is heavier than expected. This paper proposes the actual robotic system for distributing the load by adding robots or alternating the robot when a robot gets stuck. After designing the robot equipped with a telescopic mechanism, we create the autonomous flow for adding or alternating robots. It has been confirmed that the system functioned according to the designed flow throughout the object transporting experiments.

For searching a unified locomotion strategy common to different species, we focused on the reflexes and neural networks that exist in animals as control methods. In this paper, we focus on walking and undulation, explore the reflexive network controlling a snake-like robot, and compare it with that of a cat to discuss common circuits and differences due to differences in locomotion.

We aim to design an algorithm for robots to flexibly adapt and guide a target without prior information by means of inter-robot motion imitation. When performing imitation, it is necessary to compare the guidance evaluations of oneself and others. This paper proposes to investigate the validity of the terms used in the evaluation formula. We verify four local evaluation formulas that are believed to be necessary when robots evaluate the shepherding task using only their local information. The formulas include two distance metrics: distance to the target and distance to other robots. We present a scatter plot of the guidance evaluation in the absolute coordinate system and the local evaluation values. Our study confirms that it is desirable for the guiding robots to be closer to each other.

Generative imitation learning can find a reasonable policy from a limited number of successful trajectories of states and actions. However, its performance is upper-bounded by the given trajectories, which is usually given before imitation learning. Recently, reinforcement learning from human feedback (RLHF) has been receiving attention as a method to optimize the reward function with the help of human feedback. This paper proposes a human-in-the-loop generative imitation learning that learns from human demonstrations and evaluation in the systematic way. We adopt Model-Based Entropy-Regularized Imitation Learning (MB-ERIL) that estimates the reward function, the policy, and the state transition dynamics from the given demonstrations. Next, the esimated policy and state transition dynamics are used to generate a set of trajectories, which are evaluated by a human operator. Then, the reward function is modified based on RLHF. The modified reward function is further used in the forward reinforcement learning to improve the policy. The proposed method is evaluated on the benchmark task, and experimental results that our human-in-the-loop achieved the best performance.

This paper proposes a novel manipulation method for dense objects, where the distribution of the objects is manipulated based on a prediction model acquired via deep learning. The proposed method composes of two computation models. First one is a unit-type distribution transition prediction model consisting of Variational Autoencoder (VAE) and Long Sort-Term Memory (LSTM). Using this prediction model, the distribution transition of dense objects for the initial state and probe orbit can be predicted. Second one is an optimization model using Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and the prediction model. This model acquires an optimum probe orbit to achieve the desired distribution. The proposed method employing the prediction model has an advantage of rapid computing of the optimum probe orbit. The proposed method is demonstrated in experiments. After investigating the prediction accuracy with respect to the number of units, the feasibility of the distribution manipulation is validated.

Image recognition and reinforcement learning have become increasingly essential for robot control. With image data, we can obtain information on the position and shape of an object without attaching sensors to robots. Reinforcement learning can be applied directly to non-linear systems, which makes it applicable to a wide range of robotic systems. Model-based reinforcement learning (MBRL) uses data experienced by robots directly as supervisor for the neural network, resulting in high sample efficiency and shorter learning times. However, since MBRL depends on the supervised data, it could not perform well for the input image with noise. Therefore, our goal is to achieve robust control with MBRL by utilizing a spiking neural network (SNN) that would be robust against noise.

Nowadays, the number of car accidents caused by elderly people in Japan is significantly increasing. Inverse Reinforcement Learning (IRL) is a promising approach helping to reduce car accidents due to human errors in terms of accurate prediction performance on ones driving behavior. Safe driving behaviors can be imitated from positive demonstrations, and dangerous driving behaviors can also be helpful not to imitate negative demonstrations. Existing work on imitation learning form both positive and negative demonstrations has an unstable learning process due to the data conflicts between positive and inside negative demonstrations. This work focuses on the temporal locality of negativeness in negative demonstrations while most of time in negative demonstrations should be used for positive demonstrations to solve such data conflicts. This work proposes an IRL method of learning from positive and partially negative demonstrations (partial PN-IRL). By solving the data conflicts, the proposed method could reach a faster convergence and is more robust to various ratios of positive and negative demonstrations.

For robots to work in diverse and changing environments, autonomous construction of robots personalized for each user and task is necessary. We have developed an actuator module that can be reconfigured to various link structures, can carry heavy objects using a locking mechanism, and can be easily operated by direct teaching using a releasing mechanism. Given multiple target positions, modular robot configurations that satisfy these positions and minimize the required torque are automatically generated by Tree-structured Parzen Estimator, one of black-box optimization methods. We show that the automatically generated robots can perform various tasks such as moving monitors and lights.

This study aims to represent differences of the texture in pencil hardness using Intensity Segment Modulation (ISM), a sensory equivalent conversion method of vibration from high to low frequencies, which enables the presentation of delicate tactile sensations even with small transducers. We applied it to the world’s thinnest ultra-thin PZT-MEMS vibrator to a stylus-type device. The vibration waveforms of four types of pencil hardness were captured under the same conditions, and the differences in the frequency components were confirmed. We conducted discrimination tests under real writing, raw signal, ISM, and ISM cut below 100 Hz conditions. The results showed the ISM is effective in representing differences in pencil hardness.

This paper presents a load estimation method using a mechanochromic hydrogel sheet. The structural color of the gel is changed depending on the applied pressure to the gel sheet. The proposed load estimator based on the combined approaches with image features and machine learning can detect the applied load from the captured images of the gel sheet. The extracted image features of the color images of the gel sheet are superimposed on the captured initial images. By using the superimposed images as input to the machine learning system, we improve the success rate and precision of the load estimation. The experimental results show that the estimator recognizes the applied force with every 100 gf from 0 to 1,000.

The authors are developing a coated tactile sensor, ScoTacS, for robots. Firstly, we confirm the ability of the contact position estimation using prototyped two dimensional ScoTacS by measuring the delay time of the output signals experimentally. Secondly, we try to improve the accuracy of the equivalent model consiting of registance and capacitance of the ScoTacS by increasing the accuracy of approximation of the distriution of the electrical registance. We confirm the results of improvement with computer simulation using LTspice.

Surface inspection is a crucial quality control process in various industries, including automotive production lines. While digitalization and automation have improved surface inspection, human touch and sensitivity are still essential. However, detecting micro-bumps through direct touch requires extensive skill and experience. To address this, researchers developed a touch-enhancing device that amplifies tactile information and can detect small deformations, even for non-experts. The device was tested for its ability to detect micro-defects on both convex and concave surfaces, and the results showed a 2.6-fold increase in detection accuracy for concave surfaces. However, further improvements are necessary to detect convex surface deformations accurately. Future improvements to the device's structure and friction reduction could make it a valuable tool in the automotive industry's surface inspection process.

In recent years, haptics technologies that reproduce the sense of touch have attracted attention. In this study, we proposed a haptic device consisting of an electromagnet and a vibration device, aiming to present advanced tactile sensations by simultaneously presenting horizontal frictional and vertical vibration stimulations. The proposed device can present vertical and horizontal force to the finger according to the image on the display. We conducted a sensory test to determine which pattern gave the sensation of unevenness: magnetic friction only or magnetic friction with vibration stimulation. We showed that the pattern of magnetic friction with vibration stimulation may be more sensitive to the sensation of unevenness. In addition, the haptics device is presented tactile sensations simulating six types of objects using friction and vibration stimulation, and demonstrating the possibility of presenting tactile sensations.

In recent years, there has been a lot of research on deformable object manipulation. Our research group has been studying ”origami” as a manipulation of deformable objects by a multi-fingered robot hand. The shape of the paper was estimated by matching a physical model of the paper with observation point cloud. However, this method had a problem that the shape estimation became unstable when edges or a large portion of the point cloud was lost due to occlusion. In this study, we propose a method to supplement the lost point cloud with information observed by two micro cameras. In particular, a paper 3D point cloud is created from the images obtained by the micro cameras via a physical model and integrated into the main point cloud. Experimental results showed that the integration of the micro-camera point clouds with additional 3D information of the contours resulted in stable shape estimation.

The primary function of industrial robots is object handling, however, conventional handling robots are not autonomous since they require dedicated fixtures for individual tasks. A generalized autonomous handling robot requires the acquisition of a three-dimensional measurement of objects with varying optical properties, which still poses a significant challenge. To overcome this issue, this study proposes a three-dimensional measurement methodology that is independent of optical properties and a crucial prerequisite for the development of autonomous handling robots. The proposed method involves controlling the orientation of a camera mounted on the end-effector of a manipulator during handling tasks, while continuously capturing images to perform the three-dimensional measurement. The core of the proposed method is to optimize the manipulator path to enhance the accuracy of measurement and reduce the tact time. Numerical experimental results demonstrate that the proposed method can achieve three-dimensional measurement precision suitable for handling transparent objects.

Considerable research and development have been conducted in the field of manufacturing engineering automation, thereby promoting the advancement of Factory Automation. However, research and development on the automation of product disassembly are seldom conducted. Generally, products for disassembly are collected in small quantities but of varying types. Hence, the use of positioning fixtures for the automatic disassembly process is unsuitable, and disassembling by autonomous robots without fixtures is required. For instance, in screw removal processes, it is necessary to move an electric screwdriver attached to the end of the manipulator, to the screw head. To accomplish this movement, this study focused on visual servoing technology for the positioning. Nevertheless, achieving high-precision positioning with traditional visual servoing is challenging due to the lack of texture on products. To tackle this issue, this paper focuses on screw removal work and conducts disassembly using active visual servoing. A verification experiments were performed on a monitor and drum unit of a printer, successfully converging the position of the electric screwdriver to the screw head.

In recent years, demand for autonomous mobile robots has been increasing, and we plan to use markers to accurately stop robots.In this study, we focused on the fact that the accuracy of distance measurement of markers varies depending on the angle, and proposed a method to adopt as the true value the value of the camera viewed from the direction considered to have the best accuracy when reading markers with multiple cameras, and conducted an experiment. Experimental results showed that the highest accuracy was obtained at 30°. However, experiments with multiple cameras showed that it is difficult to estimate camera positions using this method because the accuracy is not stable.

Research on human age and gender estimation is currently active, and is used for security monitoring, customer statistics collection, criminal investigation. Many research have focused on methods that use clear facial information and full-body images. However, in actual usage situation, it is not always possible to obtain clear images, and estimation is difficult because some parts of the image are obstructed by obstacles or because the target’s pose makes it difficult to estimate the target. Therefore, we propose a method using full-body images with multi-task learning. By sharing information in the learning of age and gender, and by using features common to the two tasks that are useful for estimation, accuracy can be improved even for images that are difficult to estimate. The proposed model is compared with existing ones to show the effectiveness of the proposed method.

In order to reduce the constant micro-movement caused by the control noise of the mirror-driven active vision that enables high-speed magnified tracking imaging, we constructed a camera relay system using a motor equipped with a mechanical brake, and realized micro-vibration measurement of a moving object. We have verified that the proposed motor equipped with a mechanical brake has a noise reduction of about 5 times that of a conventional motor without a mechanical brake.

In this paper, we constructed an event-driven binocular vision system and proposed a time-surface-based method for calculating disparities. The event-driven binocular system is composed of two cameras that capture events, single-axis actuators that can externally adjust the baseline length of the left and right cameras, and stepper motors that can externally adjust the convergence angle. The time-surface images are obtained by applying IIR filters to the output signals of the left and right event-based cameras, with each pixel value encoding the time of the event occurrence. The proposed method obtains the disparity map by matching corresponding points in the left and right time-surface images using the block matching method. We validated the proposed method through simulation.

This study aims to automate the inspection and transport of books in libraries using mobile robots and cameras because of their ease of implementation. Therefore, this paper describes the development of book recognition technology for mobile robots using cameras. We attempted to perform template matching between the bookshelf image and the spine image using image local features ”SURF” that are robust to change of the Shooting conditions. By classifying the SURF features in the spine image, we extracted the features caused by the shape of the letters. Books could be searched with high accuracy when image local feature matching was possible.

The model-based method used for gaze estimation can obtain the 3D point-of-gaze by calculating the intersection with binocular visual axes. However, the estimated 3D point-of-gaze is unstable, and thus, error metrics are essential for evaluation. In this paper, we propose a method to evaluate the estimated 3D point-of-gaze using the similarity of scene reflection images extracted from the corneal surfaces. This method is based on a hypothesis that the similarity of the scene reflection on the binocular corneal surface is significantly reduced when the error of point-of-gaze is large. In evaluation experiments, we intentionally added an error to the coordinates of the 3D point-of-gaze, and we confirmed that the similarity of the scene reflection could be used for an error metric.

In this paper, we first propose a ”Multi-Eye Vision Hand-Arm” that can obtain visual information of around the entire robot by mounting many cameras on the robot hand and robot arm. Next, we propose a method for creating an environmental map using the ”Multi-Eye Vision Hand-Arm”. The environmental map is based on active stereo using cameras mounted on the robot. The camera’s position and posture required for stereo are obtained by forward kinematics using the robot’s joint angle sensors. The environment map is then created by modifying them with camera images. The results of an actual experiment confirmed that integrating the robot’s joint angle information and the image information from the camera can create a more accurate environmental map than estimating the camera’s position and posture based on joint angle information alone.

Several devices had been proposed to improve the musical experience. Most of them use transducers, but the problem with transducers is that the size of the device can be too large when a larger vibration is desired. Therefore, we had been studying research using air-jets. We are feeled powerful since low-frequency sounds resonating through the body at concerts and live music performances. We decided to use air-jets to stimulate the body with the sound of a bass while listen to the music. It had been found that air-jets stimulating the area around the center of the sternum enhance the sense of realism. In this study, we investigated through experiments more suitable locations of the body for the stimulation.

In this paper, we report that the pressure sensor array fabricated by the printing method is mounted on the robot gripper and the softness of the gripped object is easily determined. As a printing method, a simple method close to commercialization such as stencil printing or screen printing was used. Signals from the pressure sensor array were input to a PC through a microcomputer equipped with an ADC, and software determined the softness of the gripped object from the degree of change in pressure (inclination) and the number of sensors responding. By comparing two types of PDMS rubber with different softness, we succeeded in detecting the difference in softness.

This paper presents a novel approach to food texture enhancement that can be used on a smartphone only. Conventional food texture enhancement techniques face challenges in terms of portability and cost, but our solution offers an accessible alternative through a simple mobile application. Our focus is on enhancing the hardness of textures, a crucial factor in eating and drinking experiences. To accomplish this, we utilize a squeezing motion by the user with the smartphone to display changes in the texture of the food on the screen. Our objective is to induce tactile transfer and increase hardness. The idea was to transfer the hard tactile sensation generated in hand to the simultaneously occurring texture in the mouth, thereby changing it to a harder texture. Thus far in our investigation, we have established that the hardness of the texture is effected by the squeezing smartphone. In this study, we investigates the effect of incorporating vibration with the squeezing motion on the texture’s characteristics. Our findings indicate that the inclusion of vibration significantly enhances the thickness, density, and hardness of the texture.

Kinesthetic illusion by tendon vibration is a phenomenon in which vibrating tendons of skeletal muscles produce a sensation of body motion without actually moving. Although this phenomenon is expected to be applied to virtual reality and human interfaces, there is a lack of knowledge on how to present the time-varying kinesthetic illusion that is required for such applications. To improve this situation, we are conducting experiments to measure the transfer function for sinusoidal reciprocating motion. Preliminary data obtained so far showed that the perception was weakened when the frequency of the reciprocating motion was high, and that a phase difference of about 180 degrees was produced.

A sinusoidal transition of temperature change causes a tactile sensation and a feeling of movement on the skin surface. When the temperature rises, the palm feels an expansion sensation as if it is pushed upward, and when the temperature falls, the palm feels a suction sensation as if it is sucked downward. Since the sensation is generated only by the temperature stimulus, it solves the shortcomings of mechanical force sensing devices. We thought that the subject can change the force he or she is exerting by using this phenomenon when he or she is exerting a force. The purpose of this study is to investigate the change in the pressure value when the temperature is presented. The results showed that there was no significant change in pressure between the steady and sinusoidal temperature presentations.

We propose a self-powered stretchable artificial skin that can detect the contact position. The proposed skin is made from a lamination of an elastomer, an electret, and two electrodes. An electret is a dielectric material that can retain an electric charge on its surface. When a compressive force is applied, the positive charge induced in the electrodes is transferred through the shunt resistor connected to the two electrodes, causing a voltage drop that is used in contact detection. The internal resistance of the elastomer-side electrode is divided according to the compression position, resulting in a change of a differential voltage drop. The contact position is detected using the change. Experimental results confirmed that contact position detection with an estimation error of less than about 8 mm is possible.

The aim of this study is to enhance online communication by conveying emotions through haptic feedback. In recent years, online communication has become increasingly prevalent, but it presents a challenge in effectively transmitting emotions compared to face-to-face interactions. In this study, we developed a device utilizing pneumatic actuators to simulate a "grip" haptic feedback and investigated whether changes in emotions transmitted are influenced by factors of the haptic feedback. We designed a test device for haptic feedback and evaluated its performance. Additionally, we conducted a psychological experiment using the device. Our results indicate that the force and speed of the feedback influence participants' ratings of valence, arousal, and some of the basic six emotions experienced by the transmitter.

Soft actuators using flexible materials have been actively developed. Many of them require wiring and piping to obtain driven power. This reduces the degree of freedom of installation. In this study, we tried to develop a new actuator that enables non-contact linear motion in a space that requires an isolated environment. Light-driven molecules have attracted particular attention as a form of non-contact actuation. Conventional light-driven molecular actuators exhibit bending motion, and few of them show linear motion. In this study, we attempted to fabricate nanofibers to explore the possibility of linear motion actuators. We synthesized a light-driven polymer with cellulose as the main chain and azobenzene as the side chain. As a result, we succeeded in confirming the operation of the light-driven polymer on a thin film.

This paper discusses the design for 3D printed flexible continuum hands driven by wire systems. To perform in-hand manipulation tasks, a novel analytical design method is proposed. First, introducing a discrete model of the hand, the synergy relation of wire drive systems is derived. Next, employing the principal component analysis, the synergy relation of multiple desired postures for in-hand manipulation task is derived. Then, focusing on the isomorphism of the mathematical operation structure in the two synergy relations, a method for analytically determines the mechanism and control parameters of the hand is developed. Finally, the proposed design method is experimentally demonstrated. A 3D printed two-fingered hand driven by two wire systems is developed. By reproducing five desired postures, the developed hand succeeds in performing an in-hand manipulation task.

Wheelchair-mounted robotic arms have been developed to assist wheelchair users in their daily lives. However, it is desirable to introduce flexibility into the robotic arm to achieve more interactive tasks. To establish a design method for a flexible robotic arm, we have developed a robotic arm based on the anatomy of an ostrich’s flexible neck that can move swiftly and dexterously. In this paper, we introduce the first prototype of ostrich-inspired hyper-underactuated robot manipulator for mounting on a wheelchair. The proposed prototype was flexible yet capable of movement against gravity. This paper is the first step toward the application of a soft robotic arm with high dexterity like an ostrich’s neck to assist wheelchair users in their daily lives.

While various types of locomotion mechanisms are available for robots, this research focuses on developing a biped walking robot. The robot uses a soft, balloon-like structure for its body and incorporates buoyancy. We achieved the walking movement by a manual ON-OFF control of actuators, however, the relationship between the control signals and a gait has not been clarified yet. In this paper, we performed experiments by an automated ON-OFF control to analyze the leg joint angles and a position in the center of gravity(COG) using a motion analysis software. We also estimate a role of posture control in the gait control.

One of the various mobility aids currently being developed for the aging society and for the purpose of raising environmental awareness is a sit-only, wheeled, inverted PMV that travels on two wheels. Since this is operated by shifting the passenger's center of gravity, the passenger's movement and psychological status greatly affect the operability. However, the psychological load on passengers during acceleration has not yet been clarified. This paper describes the development of an experimental system for evaluating the psychological load on passengers using a VR driving simulator.

The rocker-bogie mechanism increases capacity of vehicles to step over steps. When it comes to climbing steps, however, the rocker-bogie vehicle depends on the length of the links and friction force between wheels and the vertical plate. Therefore, we developed the method which enables the rocker-bogie vehicle to climb a high place without friction force between wheels and the vertical plate, in which utilized elasticity. Firstly, it lifts up some links of the elasticity embedded rocker-bogie vehicle with excitation of Series Elastic Actuator(SEA). Secondly, it keeps posture of lifted links with inverted pendulum control. Then, The rocker-bogie vehicle moves forward to the step and puts some wheels there. This developed method succeeded in both simulation and real situations.

Currently, developments around robot mobility are very diverse. In disaster areas and other places where it is difficult for humans to walk, especially in places that are not flat, there has been research on methods and experiments from a variety of perspectives. Examples of previous studies include underwater movement and locations where it is difficult to obtain a stable posture, such as on stony, sandy, and soft ground. On the other hand, there are relatively few studies on snow environments such as snow damage sites. The purpose of this study is to develop a mechanism for a snow mobile robot that can move stably to investigate and supply rescue items at snow damage sites. This study investigates the mechanism by designing specialized wheels which gains both buoyancy and traction force on snow. Also, we propose a pseudo snow field that resembles physical characteristics of snow, which allows conducting experiments of snow-field mobile robots under warm environment.

Though the eccentric paddle mechanism has exhibited impressive locomotion performance on sandy terrain, its energy efficiency under loads has not been verified. In this study, we evaluated the impact of load on the propulsion capability and motion efficiency in different motion modes experimentally. The results indicated that motion mode with longer sand-contact time can produce higher efficiency.

Our crawler-type robots use direct drive motors to rotate the left and right crawlers. The response of the motors to the input differs between the crawler robot's turning motion and its straight-line motion. In order to perform accurate crawler-type robot turning movements, we analyzed the response of the crawler robot during turning. As the result, it was found that the turning motion requires compensation for turning in addition to the control for straight-line motion. It is also necessary to monitor and control the slowdown motors to prevent an increase in load during rotation. The design of current limitation is also important because of the large current flow during turning.

When disasters occur due to fire, earthquakes, terrorism, etc., it is sometimes difficult for humans to perform rescue operations. In such cases, disaster response robots are expected to be utilized to remotely operate and survey disaster sites. The disaster response robots have high traversing ability and can drive over rough terrains with proper operation. However, remote operation of robots on uneven terrains is difficult, and there is a large risk for the robot to fall over or be damaged if operated incorrectly. To avoid such a situation, automatic driving over the rough terrains is being considered to reduce operation errors. In this study, a stable automatic driving method of the robot is developed using one of the discontinuous slopes used in the standard test methods to evaluate the capability of disaster response robots.

Automated guided vehicles (AGVs) are used to transport parts and products inside factories. Recently, AGVs are often operated automatically by image processing using cameras. In this paper, we design a convolutional neural network that enables real-time image recognition even on a small computer like Raspberry Pi, and consider Path following of a mobile robot by image recognition. The training data created to train the CNN consists of line images and direction of robot movement which is determined from the position of center of gravity and gradient of the line in the image. The designed CNN model classifies line images into three classes: forward, right turn, and left turn, which are the direction of robot movement. The performance of the designed CNN model is verified by a mobile robot, and it is shown that the robot can follow straight and curve lines. the experimental results of a small mobile robot that performs line following by image recognition.