In this paper, the validation of the accuracy of a human pose estimation method in a real environment is shown. This method, proposed in the first report, uses ICP algorithm to improve the accuracy of human pose estimation using RGBD images. Some experiments to estimate climbing movement are conducted in an actual climbing gym for the accuracy verification. From the experimental results, the effectiveness of the proposed method is demonstrated by comparing with conventional methods.

Many workers suffer from low back pain. As a countermeasure, we have developed and sold a muscle suit series for waist support. In 2019, we launched a mass-produced model called “Every” that reduced the price from 500,000 yen or more to 100,000 yen and have shipped more than 20,000 units to date. The global market for power assist suits is estimated to be about 158,600 units, and it is expected to expand at an annual growth rate (CAGR) of approximately 17.1%. On the other hand, there is still no unified evaluation metric for the assistive effects of the devices, and the current situation is that each company uses its own indicators to express the assistive effects. Therefore, in this study, we focused on fatigue during work as an evaluation metric for assistive effects and attempted a quantitative expression using fingertip plethysmogram.

There are many ”unexpected” situations in our lives, and we take actions to cope with them on a daily basis. It is desirable that these actions be taken calmly, but in reality, surprise and impatience may prevent us from calmly dealing with them. One of the reasons for this is the lack of experience due to the low probability and frequency of such situations. In addition, it is impossible to know how one will react to a situation until one experiences it oneself. Therefore, we introduced VR technology, which has been developed in recent years, to develop a system that enables people to experience unexpected events, and attempted to measure people’s reactions based on heart rate variability, electroencephalograms, and questionnaires.

When humans touch objects, deformation occurs at the fingertips, stimulating mechanoreceptors that perceive tactile sensations. Therefore, there is a significant correlation between tactile sensation and deformation of the fingertip. To understand tactile sensation, it is necessary to quantify fingertip deformation and evaluate how it occurs. However, there is currently no method for quantifying fingertip deformation that takes individual differences into account. In this paper, we propose a method for measuring fingertip deformation during pressing motions using 3D scanning and analyzing deformation based on finger shape. The analysis results reveal differences in deformation based on finger shape.

In recent years, research has been conducted to automate various tasks in agriculture using robots and machinery. Harvesting incurs significant costs compared to other agricultural operations, with cabbage harvesting accounting for about 30% of the total. Therefore, cost reduction through automation is highly anticipated in this area. In this paper, we propose a method for estimating the position and orientation of cabbages to harvest them without causing damage. The position and orientation of cabbages can be determined by assuming that they are ellipsoids. In experiments comparing the position and orientation estimation using information obtained solely from an RGB-D camera and using point cloud information from a 3D LiDAR, for cabbages where the inner ball is minimally obstructed by outer leaves, both sensors yielded mostly favorable results with tilts of 20 degrees or less.

This paper describes an automatic post harvesting system of tender leafy vegetables. There is a problem that is difficult to recognize vegetables in crowd environment. The system takes RGBD images, recognizes vegetables using an instance segmentation method, calculate pose of vegetable, picks and places it at a conveyer. We chose Komatsuna as example of tender leafy vegetables. We make an annotation tool and generate datasets training the Mask R-CNN network. We show that the network recognizes komatsuna with high accuracy when we use 2-class dataset, including the whole-komatsuna class and the part-komatsuna class. We also can recognised a successful manipulation of the vegetable, grasp with good accuracy when there are few komatsuna.

In this paper, we describe a method for autonomous navigation of mobile robots using depth camera in greenhouse environments. The proposed method utilizes only depth cameras to enable autonomous movement along plant rows while avoiding obstacles. It facilitates the transition from one row to the next at junctions using junction detection. In an experimental environment simulating an actual greenhouse, the reliability of the navigation system is verified through comparative experiments with a 2D LiDAR based navigation system, which was the subject of a previous work.

Paddy fields have fill that is built around the paddy fields. Failure to control the weeds along the causeway can lead to the invasion of weeds into the rice paddies and the breeding of pests. Therefore, it is essential to control weeds at the levee slope. Weeding robots are effective in reducing the burden on farmers. But there is a problem that the weeding robot skids when traveling across a slope. For the reason, this research aimsto propose a method to suppress changes in the azimuth angle and skidding of the weeding robot. For this purpose, experiments are conducted using crawler-type outrigger arms.

The purpose of this study is to develop a sensor system for a weeding robot in order to detect the rice using an Arduino board. By wiring a stainless steel plate with the Arduino board, we can calculate the time lag. The function of the sensor system that is mounted on the weeding robot is to detect rice plants, it cannot be affected by another sensor or the robot main body. In this paper, we carried out experiment for the rice plant in the simulated paddy field. The effectiveness of the developed sensor system is verified by the experiments.

We have developed a device that can finely 3D model and measure plants for a flower and a petal in size from a few centimeters to 10 centimeters. Since image quality is particularly important in the evaluation of floral characteristics, we have developed a device that can automatically perform from 3D reconstruction to texture mapping.

In recent years, amidst the declining agricultural workforce and aging population, the burden on agricultural workers has increased, necessitating the automation of farming tasks. To address this problem, we developed an autonomous recognition and action system utilizing a car equipped with an arm for the gibberellin treatment and harvesting of grapes. Through a fusion of image-based learning methods, we significantly improved recognition accuracy even outdoors, resulting in a harvesting success rate exceeding 90% in real vineyard conditions.

In Japan, the burden on farmers is increasing due to the decline in the agricultural workforce and the aging of farmers because of the declining birthrate and aging population. In orchards, farmers currently walk around the orchard at harvest time to visually check where the fruit is growing. In this paper, we developed a system to obtain information on fruit distribution in an orchard using AR（Augmented Reality） markers that enable object detection and self-position estimation using deep learning. By using SSD300（Single Shot Detection 300） as the object detection model and segmenting the input images, we were able to create an object detection model with relatively high accuracy, although there were some false positives.

In recent years, the number of key farmers has been declining and the age structure has been shifting to older age groups, resulting in a serious labor shortage. Also, weeding in rice paddies is hard. In this paper, we propose a robot for weeding in rice paddies after drying in the middle based on a weeding method of stirring mud in “Aigamo agriculture” and a mechanism similar to that of snakes. We aim to improve the growing rice efficiency by development of the snake robot. This report describes experiment in simulated environment and rice paddies after drying in the middle, and production of the improved prototype.

Automated blade control for bulldozers have been developed. Development of those control methods requires algorithm modification and parameter tuning. Verifying those tries in real world takes much time and cost. This paper presents a novel bulldozer simulator which can be used for validating the control methods. It simulates physical characteristics of a bulldozer (e.g. engine performance, crawler torque, blade behavior). These characteristics are modeled based on the design data of the target bulldozer. The simulator also simulates interactions between a bulldozer and soil. The soil is modeled as particles. To evaluate the proposed simulator, excavation experiments are performed in the simulator and real world, respectively. The reproducibility of the simulator is evaluated by comparing those results.

Numerous research reports have been published on infrastructure inspection technologies that can detect deterioration and deformation of road surfaces, sidewalls, culverts, and other structures. Today, image AI-based technologies for diagnosis of concrete surface deterioration are in practical use and point cloud processing combined with stereo cameras and LiDAR enables 3D shape identification of cracks. However, deformation detection using advanced technology requires expert skill, and it is difficult to check all the wide-scale inspection areas under the local government's control. As a simple deformation detection method, we propose an application of the particle image velocimetry (PIV) method, which obtains vector fields from two images tracking shot by a single-lens camera. In this report, a case study will be presented in which the surface shape is estimated from the difference of movement velocity due to motion parallax using block images on computer graphics.

In recent years, the shortage of human resources at earthwork sites has been accelerating due to industrial accidents and the declining birthrate and aging population. Against this background, Our lab is developing a CPS platform called ”ROS2-TMS for Construction” as a system that simultaneously improves the efficiency and safety of earthwork work. This paper introduces a newly developed task management mechanism as a continuation of the ROS2-TMS for Construction development. The task management mechanism is a mechanism for operating actual construction machinery based on scenarios called task sequences given to ROS2-TMS for Construction.In this time, I adopted the Behavior Tree as the task scheduler and incorporated it into ROS2-TMS for Construction with some extensions according to the ROS2 specification. An overview of the extensions of Behavior Tree we adopted to task management mechanism of ROS2-TMS for Construction will also be explained in this paper.

This paper presents an interface of ROS2-TMS for Construction, a cyber-physical system designed to improve the efficiency and safety in earthwork, for an OPERA (Open Platform for Earthwork with Robotics and Autonomy) compatible backhoe. We focused on the excavation and loading operations of earth and sand and confirmed that the proposed system performs these operations appropriately on an actual OPERA-compatible backhoe.

This study develops a deploying method for a maintenance robot to a Bridge Maintenance Rail Structure. Attaching the robot to the rail structure requires high positioning accuracy. However, it is difficult to achieve such positioning accuracy in a hard-to-access location such as under a bridge. This method functionally separates the maintenance robot into a guide unit, a working unit, and a transport unit. This method replaces the task to be performed under a bridge for robot deployment with the docking of the working unit and the transport unit. This task replacement greatly increases the the tolerance of positioning for deploying the maintenance robot to the rail structure.

Semi-autonomous control of a hydraulic excavator at disaster site and construction site are required from the safety and efficiency points of view, and some autonomous controls have been reported so far. To realize efficient semi-autonomous excavation, the operator must be able to determine the information necessary for operational intervention on the operating device. However, conventional methods have few means of presenting information, and thus have problems in adapting to various functions, such as selecting multiple tasks, and in ensuring the safety of information presentation. In this study, a semi-autonomous excavation system is designed by introducing a virtual spring mass damper system to the joystick as a control device, thereby increasing the number of information presentation methods and enabling the presentation of multiple work directions and safe presentation of external environmental forces. The effectiveness of the proposed method is evaluated through simulation and experimental verification.

Concrete structures are composed of three-dimensional surfaces, making them complex moving environments. We constructed a stability model and developed a novel robot on the floor and wall surfaces required for the robot to move on a three-dimensional surfaces. The proposed robot can move not only by utilizing the thrust generated by its propeller but also by deflecting the thrust in the appropriate direction. Based on this deflecting thrust, we clarified the conditions under which the robot does not slide or fall on the 3D surface and developed a theoretical stability model. In addition, experiments were conducted to verify whether the robot could move from a floor surface to a wall surface. As a result, it was confirmed that the robot can move between two surfaces.

It is important that the logical and economical control of air conditioning of a large building with heat source using the combined system of BEMS (Building Energy Management System) and AI. In this study, we propose a new estimating method of heat source for building air conditioning to use the AI system. The method easily estimates the heat demand of buildings through computational fluid dynamics and a basic unsteady heat conductivity model. The fundamental estimating model and implementation factor are determined by actual data of the building B of Fukuoka Institute of Technology. The applicability of this method to another building is investigated, which are not obvious their thermodynamic and structural material data. Consequently, the present method can reasonably estimate heat demand of a building.

Unmanned construction is a method of construction in which heavy machinery is operated remotely, preventing secondary disasters at disaster sites. Compared to on-board operation, remote operation reduces work efficiency by about 50%, so appropriate video presentation to the operator is necessary. A drone-based video presentation method has been proposed to improve work efficiency, and it has been suggested that work efficiency can be improved by arbitrary control of video. However, viewpoint manipulation interfaces have not been sufficiently studied. Therefore, in this study, we analyzed the effects of different viewpoint control interfaces on teleoperation efficiency.

The purpose of this research is to realize the predictive failure diagnosis of hydraulic actuators of dust trucks using a recurrent neural network - autoencoder (RNN-AE). In this paper, the performance of principal component analysis (PCA) and RNN-AE are compared as methods for extracting time-domain and frequency-domain features from the acceleration data of a hydraulic actuator, mainly in terms of compression performance. As a result, a total of 15 time-domain and frequency-domain features were extracted from the side acceleration data of the vehicle body, and it was found that the compression performance of PCA and RNN-AE was almost equal for the 15-dimensional features. A future work will be to apply it to calculating the degree of abnormality from the compressed feature values to realize a predictive failure diagnosis of hydraulic actuators.

Our research group has been developing a flexible and buckling-preventing SCARA robot as a confined-space exploration robot that can be mounted on a quadruped robot. However, the constant restoring torque joint used in the proposed mechanism has a non-functional zone where no restoring torque is generated near the linear state of the joint, which is a factor that reduces stability. Therefore, in this study, we investigated the improvement of the mechanism and its implementation in a quadruped robot. To remove the non-functional zone in the joint mechanism, a joint mechanism using magnetic springs was implemented in an actual machine, and the partial removal of the non-functional zone and the improvement of stability and convergence of joint motion were confirmed through experiments on the actual machine. In addition, we selected composite materials for weight reduction and discussed and designed a reel mechanism to accommodate the proposed mechanism in a compact size.

Autonomous mobile robots (AMRs) ordinally move on level grounds. However, several types of tough AMRs have been designed and developed for locomotion on uneven grounds. They are employed in agricultural robots, disaster robots, and robots used for hazardous applications. Especially, semi-active damping elements are one of the smart structures and potential solutions for highly stable locomotive robots. In this study, we propose a vibration control unit with a Magnetorheological Fluid (MRF)-based semi-active universal joint (SAUJ) for stable delivery equipment that can be attached to an AMR moving on uneven ground. The basic structure and controller of the SAUJ are proposed. In addition, its performance is experimentally evaluated on an AMR moving on uneven ground. According to the experimental results, vibration control of the SAUJ successfully decreased the perturbation of a pendulum swing due to the acceleration of the AMR.

This paper is concerned with an advanced automatic pouring control system with 6-axis vertical articulated robot. In the previous studies, many automatic pouring control approaches have been proposed for applying automatic pouring machines used in casting industry. In order to diversify the pouring control approaches into other fields, we develop the automatic pouring control system installed into the 6-axis vertical articulated robot in this study. The pouring flow rate control and filling weight control are integrated systematically in the proposed approach. The efficacy of the proposed pouring control approach was verified by the experiments with the vertical articulated robot.

This paper contributes to a training simulator for acquiring efficiently pouring skill in a foundry. The real pouring work is dangerous environment for the worker because treatment of the high temperature molten metal. On the other hand, the skilled operation is required for pouring precisely the molten metal into the mold. Therefore, we develop the training simulator for acquiring safely the skill of pouring work can be indicated quantitatively with the graphic representation for acquiring efficiently the skill of pouring work. The efficacy of the developed training simulator was verified by the traininig experiments.

In this study, we propose the self-localization system using landmarks recognition for autonomous transporter. The laser range finder (LRF) and the RGB-D camera are used to obtain the locations of the landmark’s centroid in the cartesian coordinates. The derived landmark locations are stored on the computer and used for the self-localization. In the proposed self-localization, the local location of the transporter can be estimated using the principle of trilateration with the locations of landmarks and the transporter based on the sensor coordinate. In this approach, the values of all sides and angles in the triangle which consists of the locations of landmarks and transporter can be derived from the geometric relationship. And the global transporter location can be obtained by comparing the difference of the triangle of previous sampling time. The efficacy of the proposed self-localization system is verified by the experiments with the transporter.

This paper is concerned with an advanced admittance control to perform efficiently cooperative works between a human and a robot. An assembly work of heavey parts is one of hard physical works. Therefore, it is required to utilize a collaborative robot for assisting the worker. However, it is difficult to manipulate directly the robot at will. In this study, we propose the admittance control installing the virtual model with the bounded viscous resistance for improving the operability of the direct manipulation to transport the heavy object by the cooperative work of human and robot. Improving the operability by the proposed approach can be analyzed by the frequency analysis. And, the efficacy of the proposed approach is verified by the experiments with the collaborative robot.

Machining systems using industrial robots is expected, but existing systems face challenges in expanding workspace and supplying materials necessary for processing. Therefore, we have proposed Modularized-Structure and Multiple-points Simultaneous Machining System, MS-cubic. In this paper, we have proposed an approach to realize 2D movement of robot arms using MS-cubic. As a result, we achieve drilling at positions higher than the work range of the robot arm. Additionally, vibration at the end effector of the robot arm decrease during drilling operations because the scalability of MS-cubic enables the robot arm to maintain a posture and position conducive to machining.

In current polishing robots, automation of the polishing operation by force control of the force measured by force sensors is being promoted as the mainstream. However, skilled technicians use not only the sensation of the force applied during polishing, but also the sound information generated during the polishing process. In this study, we investigated the possibility of adding sound information to polishing robots to expand their functions and improve their control accuracy. As a result of frequency analysis and measurement of the effective value of sound, it was found that it is possible to determine the wear condition of the tool and the optimal pressing force from the change in the effective value of sound information.

This study proposes a playback trajectory correction method for peg-in-hole task on a misaligned workpiece by using our developed parallel wire-type teaching and playback device. This device combines our developed parallel wire-type teaching device (PAWTED) and a remote center compliance device. The proposed method allows the playback trajectory to be corrected to match the misalignment of the workpiece considering the compliance of the robot, thereby increasing the speed of robot. Experiments simulating peg-in-hole task were conducted on a workpiece with multiple holes, and it was possible to reduce the work time by 11.3% by correcting the trajectory according to the misalignment of the workpiece and increasing the speed of the robot.

In order to use lithium-ion batteries safely, it is necessary to understand the battery’s lifespan. The purpose of this study is to use data obtained from charging and discharging batteries to predict their lifespan. We conducted an experiment in which we classified the data into three categories: short life, medium life, and long life, and trained each LSTM to make predictions.

In this paper, the authors prototyped a learning and development framework that seamlessly links ROS learning and robot development using ROS for students and engineers whose main purpose is mechanism development. The developed learning material is based on a commercially available robot kit. The process of adding peripherals is simplified from a hardware perspective by the use of an IoT sensor system, and from a software perspective by the preparation of code templates and procedures. For post-learning development, a control system that repackages the learning materials is used, thus providing a learning and development framework in which the learning outcomes can be utilized. The usability of this framework was confirmed through the development of a small robot for paddy fields.

In recent years, the aging of the population has led to expectations for automated driving systems for automobiles.However,automated vehicles are expensive and d ifficult to use.As a solution to this ploblem,the development of retrofitted automatic driving systems(hereinafter referred to as “portable automatic driving systems”)is underway.However,currently developed portable automatic driving systems have only limited elements such as steering wheel operation or gas pedal and brake operation,and there are still elements that require human operation.To solve these problems ,we are researching and developing a portable automatic driving system that enables both steering and gas pedal,and brake operations with a simple retrofitted mechanism.We also aimed to develop a highly versatile mechanism that can be adapted to various shapes.The usefulness of these mechanisms was examined by testing them on actual mechines.

Since 1988, the Department of Electronic Control System Engineering at National Institute of Technology, Numazu College has conducted development of Micro Intelligent Robot System “ MIRS ”as a Project-Based Learning (PBL) into the curriculum for the education of system development. The purpose of this project is to provide a place to learn ”the process of system development,” ”education as creative education,” and ”social implementation experiments” through robot development. This paper reports on the MIRS development and education program undertaken from FY2017 to FY2023, reviews past efforts based on man-hour analysis, and discusses issues for the next generation. As a result, it was effective to set social implementation experiments as an achievement goal for the robot to be developed. However, it also became clear that it is necessary to plan lessons and provide guidance so that students do not develop devices with only social implementation in mind.

This study deals with a situation in which the military forces of one country are guarding a base and combatants of the other country attempt to invade the base. In this situation, a formulation of precautionary measures regarding the principle of distinction is proposed. In the distinction between civilians and combatants using a DNN based on camera images, the certainty of the distinction is important. In previous studies, “effort” was proposed as a criterion of how much effort was made to obtain this certainty, but it was expressed as a constant value and did not change depending on the situation. In this study, as a feasible criterion, it is proposed to vary this “effort” according to the distance from the combatant to the base. Numerical simulations show that the proposed criterion can be used to prevent the invasion of combatants from one side while preventing accidental attacks on civilians.

This study introduces a Python-based motion experimental interface for industrial dual-arm robots (This study focuses on ABB’s YuMi). The interface facilitates robot motion verification and experimentation by providing an intuitive graphical user interface (GUI) for specifying robot motions. Users can input parameters such as end-effector coordinates, joint angles, and velocities without requiring knowledge of ABB robot’s dedicated language, RAPID. The interface establishes a connection between Python programs and RAPID, enabling combined robot control. This paper discusses the design, implementation, and experimental results of the interface in both on simulation and on physical robots. The proposed interface aims to improve the efficiency of motion planning, verification and experimentation for dual-arm robots, which are increasingly being applied to complex tasks in industrial settings.

In recent years, factory automation has been promoted due to labor shortages.To solve this problem, labor saving and efficiency improvement by introducing collaborative mobile manipulators is attracting attention. The purpose of this study is to develop a collaborative mobile manipulator that emphasizes the vertical working range. A system is integrated using ROS to develop an overall system that controls a collaborative dual-arm robot, an omnidirectional movement mechanism, a lifting mechanism, and sensor system consisting of cameras and LiDAR, and has a function generates movements corresponding to AR markers as landmarks. The results show that developed collaborative mobile manipulator can perform a wide range of tasks and it demonstrated the detection cargo and accomplished cargo transportation tasks.

Marine pollution caused by the discharge of waste has become a global issue. In this study, a system to facilitate the collection of drifted debris on the coast by enabling easy understanding of its current state was developed. The proposed system could quantify of drifted debris by multiple user, using garbage detection AI through uploading images from mobile devices such as smartphones to cloud storage. The developed system was verified by 92 users during an actual coastal, and it made it possible to collect 317 debris images and detect 956 debris items.

In this study, we developed an autonomous mobile robot platform that can be used in various environments. The platform enables the development of robots with additional functions to the mobile robot platform without having to build a robot with the function of mobility from scratch. As an example of application of the mobile robot platform, we developed an autonomous mobile robot for transporting parts in a factory. Experiments were conducted in a variety of environments, including indoor, outdoor, and factory environments, and the results were evaluated.

In an environment such as a disaster site, debris is likely to be scattered about, and a mechanism that can move over uneven terrain environments is required. In addition, it is difficult to determine the exact position of the robot by remote control, and a moving mechanism that can overcome uneven terrain without control is needed. We have developed a moving mechanism that accomplishes these tasks. The name of the mechanism is ”Odd Bogey Mechanism. In this study, we developed a mechanism for decommissioning work at a nuclear power plant, which is one of the disaster environments, and mounted it on a robot to evaluate its performance.

Imitation learning enables robots to autonomously perform complex tasks that are difficult to accomplish with conventional technologies, and the range of applications continues to expand every day. However, generalization performance is highly dependent on the quality of the collected data, and furthermore, variations in motion are limited to teachable tasks. Therefore, a robot system that can easily collect a variety of teleoperation data with high quality is essential. In this paper, we develop a dual-arm mobile manipulator and an intuitive tele-operation system that can be used in narrow spaces and for high/low level tasks, and confirm that it is capable of teaching multiple daily life tasks.

A robot that can perform various tasks in various environments and operate flexibly and stably even in unexpected disturbances is desired. In this paper, we proposed a motion generation model using 4ch bilateral control for tasks involving contact with environment. A hierarchical recurrent neural network structure was used to simultaneously control position and force based on visual information. An evaluation was conducted on the USB connector insertion operation using the proposed motion generation model, demonstrating a success rate of 80% against a positional deviation of ± 5mm in all directions. In conclusion, it was confirmed that the proposed motion generation model can integrally predict/control position and force.

In object detection using verbal instructions, ambiguous instructions may fail to identify objects. In this study, we developed an interactive object detection system that disambiguates instructions by obtaining the most useful appearance features for identifying the target object through dialogue, based on the results of appearance comparisons of multiple detected objects. Through experiments on the detection of commodities, we have confirmed that the proposed method can detect the indicated object even with fewer questions.

One of the challenging issues in robotics is dealing with uncertainty. In this research, we propose a method for robust motion generation in unlearned environments, including extrapolation, by introducing Bayesians to conventional deterministic RNNs. The proposed method learns a probabilistic model from training data and generates situation-specific behaviors from the learned probabilistic model. We confirmed that the proposed method can robustly execute tasks even in unlearned environments by using a robot simulator. Furthermore, we show that the Recurrent Dropout can be used to reduce the dependence of the initial weights of neurons on generalization performance.

It is distress for a patient on the bed to pick an object up that is away from the bed. A robot may solve such a problem, that is, if the position of the object is known, a robot can pick it up and deliver it to the patient. However, there are usually lots of objects in the room, and their positions are unknown. This means that the estimation of the position of the object that the patient would like to pick up is required for picking by a robot. In this research, the authors proposed a method to estimate the position of an object by taking pictures with an AR marker. This AR marker is attached to the wall, so its position is previously known; it is possible to be used as a reference position. Experiments were conducted, and the precision of the position estimation were evaluated.

Energy Sustainable Robot (ESR), capable of moving and feeding using environmental energy sources,utilizing leaf biomass is proposed. We aim to develop and integrate components for mobility, collection, combustion with thermoelectric generation, and power storage. Deep learning for leaf detection and reinforcement learning for approach behaviors are crucial for autonomous operation in uncertain environments. Our study achieves leaf detection using SSD and acquires leaf approach behavior via Dyna-Q. Both leaf detection capability and approachability are verified by experiments.

This paper evaluates the power consumption of an autonomous mobile sensing robot that observes a fixed-point environment while moving through multiple locations on the water, when the robot transforms into an efficient form in each case of straight ahead and when it maintains a fixed point. In previous studies, we have measured the power consumptionofawatermobilerobotwithapoint-symmetric(X-shaped)chassisduringcontinuousstationarymissions and during traveling missions that alternate between stationary and straight-line motions, and have studied efficient motion strategies from a software perspective. Experiments measuring the power consumption during traveling missions showed that the power consumption during straight-line motion was higher than that during fixed-point maintenance missions. Therefore, a transform mechanism was developed to achieve an efficient thruster configuration and shape stability during both fixed-point and straight-line tasks, and the effectiveness of the transform mechanism was verified in real field experiments.

Underground inspection systems in closed tunnels are expected to be replaced by mobile robots. In this study, an underground inspection robot system was developed, and field tests were conducted to evaluate the system and identify issues. In the field test, it was confirmed that automatic traveling, self-positioning, and shooting in the dark were possible. It was also found that the cause of failure in automatic traveling was a breakdown in self-localization.