In recent years, ICT in agriculture has been recommended. Because raspberries have different harvesting periods for each fruit and harvesting work is prolonged, I think that it is necessary to mechanize the harvesting work from the viewpoint of profitability. The purpose of this research is to identify the raspberry fruit in the harvesting period and specify the position in the camera image. Ultimately, the goal is to construct a camera system that designs the optical filter by directly using the spectroscopic analysis result derived from the functional substance, and identifies harvestable fruits that contain a large amount of functional substances. In this paper, we report on the consideration of fruit discrimination by bandpass filter based on spectroscopic analysis result.

In this paper, the authors describe the development of a small robot for measuring chemical ingredients and environment information such as nitrogen, pH and electric conductivity etc. to know distributions of these in paddy fields. Developed robot is radio control device that has floaters in front and behind of its body and travels paddy fields by two wheels. The robot inserts the sensor in the ground and send the measured data by radio communication. Operation check is done in the cistern and the paddy field.

Agriculture field has been improving to meet humankind important need, which is food. Farming went through different stages from traditional ways to mechanized ways nowadays and even irrigation, which is the most important part of farming. Currently, the proportion of old people involved in agriculture is increasing. In order to meet the increase in population number around the world we have to find agricultural solutions, which are more productive. Because of this, researchers and engineers started to design and implement agricultural robots, which is an extension of agricultural machinery revolution. In regards to tilling works, which requires considerable physical strength to operate tractors and this, could be a problem for old people. Therefore, our purpose in this paper is to present a small tillage robot for home gardens.

Rotary speed have to accurate control to realize high operability in electric tiller. Tilling reaction torque is periodic. Thus, repetitive control is suitable for speed control in electric tiller. However, conventional repetitive control does not correspond to period variation of tilling reaction torque due to variation of rotary speed. Moreover, tilling reaction torque has jitter due to variation of ground condition. This paper proposes repetitive control system which compensates tilling reaction torque with jitter. Proposed method uses angle of loader because tilling reaction torque depends on angle of loader. This paper confirms the effectiveness by experimental results.

Recently, the labor shortage is becoming more serious because of decrease and aging of farmers in Japan. As a solution therefore, the automation of agricultural robots are trying to expand the scale of management and labor saving. This paper describes a study of an automated parallel running robot tractor for automatic harvesting. The goal is to run a harvester and tractor towing the containers to hold harvested crops in parallel. This study uses a high accurate self-localization method using RTK-GNSS. However, since the output frequency of RTK-GNSS used this time is relatively low, UKF was applied to estimate in high frequency with IMU. Also, the position and attitude information of harvester was obtained by wireless communication. We could not obtain sufficient accuracy because of lack of accuracy of UKF in experiments. Sufficient accuracy of UKF was obtained in the experiment conducted thereafter.

In Japanese mountainous region, there are many levee slopes, in which weeds are grown and effective vegetation are needed. Conventionally, string trimmers are used to conserve the vegetation. At the same time, however, it is true that some of the users, especially of the elders, have an accident by them. Therefore, it is difficult to maintain. The Japanese government, accordingly, has been promoting the automation of agricultural machine to prevent such accidents. This research describes wheels for weed cutter robot that has a mechanism for changing camber angle on steep levee slopes. The usefulness of it, moreover, was evaluated by simulation.

In recent years, drones are used reconnaissance tasks in dangerous places where people cannot enter. Drones are also used in spaces where it is difficult for people to work directly, such as checking the exterior walls of buildings and piping inside the building. However, there are problems that the posture becomes unstable due to the air current emitted by itself in the vicinity of a structure such as a wall and the rotor rotates at high speed, causing breakage. In this paper, we propose a drone with a small-sized contra-rotating tilt rotor covered with a revolvable spherical exterior as a protection of the rotor to secure safety of surrounding people and rolling wheel to move on the ground. Development of a prototype and operation test are shown.

Flexible continuum robot is useful for search and rescue operation in narrow spaces. We have proposed air jet type Active Scope Camera floating the head of the robot by air jet. The air float hose type robot can move over higher steps and change the moving direction easily. In this paper we suggest how to change the air jet direction around two axes, and how to choose the geometric parameter.

Recently, the quadcopter has attracted much attention in the various industrial fields. Our research plans to use the quadcopter in the forest industry. The quadcopter is required to have high flight capability because there are a lot of obstacles in the forests. However, typical quadcopter has four rotors, that is, only four inputs. Therefore, the typical quadcopter is a kind of under-actuated systems. In our previous research, we developed the tilt angle control unit, which can change tilt angle of each rotor with respect to the body of the quadcopter. By equipping with four units to one quadcopter, control inputs are eight and the quadcopter becomes a fully-actuated system. In this paper, we confirm whether the quadcopter is controllable by computer simulations. Next, we show the hardware and software configurations of the quadcopter. Finally, we consider several assignments through the experimental results.

This paper presents a semi-autonomous collision avoidance flight using two-dimesnional laser range finder with mirrors. The sensor used in this research scans over horizontal and vertical directions simultaneously. We develop obstacle detection and height estimation system for inspection under bridge by using this sensor. The hough transform is used for the obstacle detection in the flight direction. For the height estimation under bridge, we develop a method to estimate the relative distance between the UAV and bridge even in the case when the surface of the bridge is rough. Flight experiment is carried out to verify the validity of developed system.

This paper proposes aerial torsional manipulation by using a quad tilt-rotor UAV. Since the conventional multirotor UAV is an underactuated system and has low mobility, possible aerial manipulation is restricted. On the other hand, the quad tilt-rotor UAV has features of producing high torque around yaw axis and achieving independent position and attitude control by adding a mechanism to change the thrust direction with respect to the base body. In this paper, we develop an angular gripper to realize aerial torsional manipulation which is mounted on the quad tilt-rotor UAV. Finally, we aim to realize removing a light bulb using the quad tilt-rotor UAV as an application of the aerial torsional manipulation by virtue of this feature.

In non-GPS environment such as in a valley of large buildings, it is essentially difficult to automatically control a drone. Although some SLAM-based control techniques have been recently proposed, it will take a while to build a precise map as a preparation to operate the drone. In this report, in order to instantaneously operate the drone, we propose a precise position control method using a guidance light on a drone station and a ring-shaped sensor mounted under the drone.

This study develops a pumpkin peeling robot with a compliance mechanism. The robot consists of three servo motors, which are a roll axis motor, a yaw axis motor and a motor for chucking and rotating a pumpkin. And a very simple compliance mechanism without using any force sensor is realized by a contact switch and a spring element between the peeler and the robot arm. The compliance mechanism can keep the peeler contacting with the pumpkin surface with strong roughness. This paper shows the structure of the robot arm, the pumpkin chuck and the simple compliance mechanism, and explains the control system and procedure of the peeling operation in detail.

The purpose of this study is to develop a motion control method that enables autonomous running and a sensor system to detect rice plants of weeding robot in paddy field. The motion control method consists running along rice plants rows and turning. The sensor system consists of six capacitive touch sensors to detect the rice plants. In this paper, we carried out running experiments using imitations of the rice plant in a paddy field. The effectiveness of the developed motion control method and sensor system is verified by the experiments.

Recently automation and labor saving in agriculture have been required. However, mechanization and robots for growing fruits have not been advanced. In this paper, a method of detecting fruits and automated harvesting by a robot arm is proposed. To detect the position of fruits, a highly fast and accurate method with Single Shot Multibox Detector is used. To detect the three-dimensional position, a stereo camera is used. After calculating the angles of the joints at the detected position by inverse kinematics, the robot arm is moved to the position of the target fruit. The robot harvests it by twisting the axis of the hand. From the experimental results, more than 90% of fruits were detected. Also the robot could harvest a fruit in 16 seconds.

In recent years, the agricultural work population is aging and decreasing, and in addition, agriculture successors are also decreasing in Japan. The reasons of them are probably hard work and low income. As the solution of this problem we focus on raspberry harvest automation. Raspberry fruit contains a large amount of compounds with high health maintenance effect, so it will be a high added value product. This will save the declination of agriculture in Japan. With that back ground, we propose to automate cultivation and harvesting of raspberries. In this paper, especially, we describe our proposed automated-harvesting system.

We have developed a weeding robot(Aigamo Robot). This robot works automatically using Global Navigation Satellite System (GNSS). Previous independent positioning measure included several meter differences and dispersion of the positions and directions. Therefore, in this study, we implement the Real Time Kinematic-GNSS (RTK-GNSS) on the right and left side of the robot in order to reduce these errors. This method can obtain the position and direction more acculately when the robot is not only in operation, but also under suspending. Moreover, this method can suppress the errors within about 10 cm to the south-southeast when the path tracking.

GPS is used for identifying the self-location of the robot outdoors. GPS cannot be used when the radio wave for GPS signals from satellites are shut off by mountains and buildings. Therefore, we have developed an instrument that measures the position of a robot instead of GPS. The instrument is named a tracking laser rangefinder. We manufactures a prototype autonomous mobile mowing robot and carried out performance test. The robot could trace the straight line paths, but slightly derailed the circle paths.

Low-speed-descent objects such as an object descending with a parachute are widely used in the aerospace field. They are used for ensuring safety or making the success of their missions but have some disadvantages. Mid-air retrieval of the low-speed-descent objects by aircraft is discussed to resolve their problems in this report. This paper describes fixed-wing UAVs' control theories of terminal approaching and shock response control after retrieval. Updating final state control and model predictive control are used as the control method of the approach. Frequency-shaped LQR is applied to the flight control after mid-air retrieval. The fundamental method for achieving the mid-air retrieval is constructed with these methods and its effectiveness is verified by simulations.

We aim to search by automatic piloting of known areas (indoor) under GPS-denied environment. Specifically, the wall surface of a room without obstacles is set as a course, and it is searched by autopilot. The sensors used for automatic piloting were selected as follows. Since the objective of this research is automatic piloting in the unsearched area, it is not appropriate to self position estimation by giving the external environment in advance, such as AR marker and motion capture. Next, Visual SLAM is not appropriate for indoor use because data cannot be obtained in dark places, and it is difficult to recognize walls with little color change. Therefore, we proposed a method to measure and avoid the distance of obstacles using the laser range finder.

In recent years, functional disorder and road sinking caused by corrosion of sewer pipe have become serious problems in our country. A lot of practical surveys in the sewer pipe are currently carried out by using self-traveling vehicles with camera. However, this kind of survey method is expensive and requires arrangement in advance. Therefore, we focused on an investigation method by flying quadrotor helicopter what is called “drone”. In this research, we study simulation experiment of AR.Drone using ROS. First, we have constructed a simulation environment using ROS. Next, we have created the flight program. Finally, we have extracted feature points in sewer pipes.

This paper describes development of 3-DoF flapping-wing robot with variable amplitude link mechanism to control lift and thrust forces. The variable amplitude link mechanism has lever crank mechanism driven by a BLDC motor and a linear acutuator to control amplitude of flapping angle. The robot has also 2 DC motors for feathering and lead-lag motion. Some experiments of measurement of force-torque revealed effects of each wing motion. We found that flapping amplitude difference between left and right wings occurs roll and yaw moment.

This paper presents a development of an aerial robot for the use of close inspection of vertical walls and columns made of steel. The proposed robot has magnetic clinging devices in front and rear ends of the airframe, each which consists of an EPM (electro-permanent magnet) with a rotation joint. Using these devices, the robot approaching to the target structure can cling to the structure, and conduct inspection missions such as image inspection and vibration monitoring. In this study, the performance of the clinging device is experimentally examined on test pieces with different curvatures, considering various situations of approach with different offsets and angles, and the allowable limits of the collision speed are discussed.

This paper proposes a method to measure wind direction and speed by a multi-copter UAV. The main purpose is to measure ocean surface wind, which is difficult to measure with conventional methods. The proposed method estimates the wind direction and speed in real time using the attitude of the UAV, based on the force equilibrium on the UAV. Pitot tubes are also attached to the vehicle. The crosswind performance of pitot tubes are evaluated at wind tunnel. The prototype UAV was developed and the performance of the proposed method was evaluated though sea trials.

While social robots become ever more popular, novel modeling and control methodologies are sought toward optimizing their engagement. This work considers a NAO robot playing games with children. A novel feedback control is proposed using a resultant sentence, induced by crowd-computing techniques, as Reference to be compared with a computer-vision induced sentence toward driving a linguistic controller. Preliminary application results have been encouraging.

The effectiveness of human-robot interaction depends on the capacity of a robot to modulate its behaviour according to human emotion(s). This work presents preliminary results regarding a Behaviour Modulation System (BeMoSys) implementable on a social robot with a capacity for emotional speech recognition based on signal-processing followed by machine-learning techniques.

There is a breath of new experiments recently being carried out on the introduction of robotics in many educational and therapeutic related activities, ranging from children with the autistic spectrum condition (ASC) relation to anthropomorphic robots to the effect of robotic programming in the cognitive capabilities of healthy and diagnosed children. Here we review the way these experiments measured the effects of robotics insertion in education.

This research describes an assist robot which harvests citrus fruits. The robot is required to have a function of autonomously moving between a worker and a monorail which carries fruits to the foot of the mountain. However, in general, there is no standards in each orchard, the situation is different in each place. Therefore, the robot has a function that estimates the position of the worker. Because no special qualification for using XBee is necessary, the XBee is used. In case that a transmitter is attached to the worker, it is verified that XBee with a parabolic reflector estimates the position of the transmitter.

The purpose of this research is to develop the autonomous lawn mower controlled by using GPS. We aim to construct a system that mounts a control system on existing lawn mowers and automatically harvests turf of a predetermined range. In order to operate the robot, it is necessary to measure its own position. There are various positioning methods for self-position estimation. we can't use distance sensors so we develop the robot using satellite positioning and inertial positioning. The lawn mower runs along the target line by controlling steering. Its speed is constant. The target angle of steering is determined from the target line, the current position information from the GPS and the current direction of the lawn mower. The steering angle is controlled by PID control.

Accidents in farm work by brush cutter have reported annually. In this study, to aim at development autonomous mowing system, the method for mowing system is proposed to implement autonomous running by range specification and verify the effect of the method on actual environment. This method includes target track generation that generate the zigzag line in specifying range, moving method limited rectilinear propagation and pivot turn, and self-position estimation separated for only during rectilinear propagation and for only during pivot turn. Finally, the demonstration experiment is performed on actual environment.

To increase payload and efficiency, stabilizer-less unmanned helicopters are highly demanded. The purpose of our research is to develop flight controllers for the stabilizer-less unmanned helicopter. We focused on attitude controllers applying dynamic inversion method to cancel undesired coupling properties and to realize desired response. In this paper, we applied hierarchical approach of dynamic inversion method to design attitude controllers based on the original dynamical model of horizontal motion of the stabilizer-less unmanned helicopter, a low frequency model, and a steady flapping model, which was called as Hohenemser model. The delay due to the servo motors was not usually included to design controllers not to place too much load on them. However, it is possible that servo delay has significant influence on the flight quality. Frequency analysis of response to wind disturbance with servo delay was carried out. Results show that the flight controller designed based on Hohenemser model is not appropriate because the servo delay induces undesired oscillation of the main rotor's flap angle.

A general flying automobile has two type mechanism. One is based on fixed-wing aircraft, and the other is based on rotary wing aircraft. In this study, we aim to develop a rotary wing aircraft model. Based on a rotary wing aircraft, we can make it smaller size than a fixed-wing aircraft model. Additionally, a rotary wing aircraft can take off and land vertically. Therefore, there is the merit that we can use it on the general road without needing a special runway. On the other hand, there are some problems to solve. For example, we have to secure enough thrust to let a heavy body fly, and have to consider the influence by the change of a center of gravity accompanying weights of occupants. In order to solve these problem, we consider control characteristics on the experimental aircraft.

Since a conventional multi-rotor UAV is an underactuated system due to the fixed thrust direction with respect to the base body, its attitude changes during translational movement and its position and attitude cannot be controlled independently. Therefore, a quad tilt-rotor UAV has been developed which is able to fly with control of six degrees of freedom by adding additional actuators to change the thrust direction. However, since the number of actuators increases, the control system becomes complicated. It induces difficulty of gain adjustment for stable flight. In this research, focusing on attitude control, we aim to develop a control system that does not require trial and error gain adjustment for a quad tilt-rotor UAV, and evaluate its validity with numerical simulation and flight experiment.

We add a clustering method to Particle Filter on Epsiode (PFoE), which is a teach-and-replay method for a robot. The clustering method is used for classification of similar sensor data on the memory sequence obtained in the teaching phase. The improved PFoE is more robust toward the noise of motion and sensor information than the previous PFoE. That is because the classification result gives the information which phase in the replaying task the robot performs. The difference between the proposed and previous methods are compared by a task with an actual robot.

This paper proposes a policy transfer method of a reinforcement learning agent for suitable learning in unknown or dynamic environments based on a spreading activation model in the cognitive psychology.The reinforcement learning agent saves policies learned in various environments and learns flexibly by partially using suitable policy according to the environment. In the proposed method, an undirected graph is created between policies, and the network is constructed by them. The agent updates the activate value that policy has according to the environment while repeating processes of recall, activation, spreading, attenuation and learns based on the network. Agent uses this network in transfer learning. Experimental simulations comparing the proposed method with several existing methods are conducted to confirm the usefulness of the proposed method. Simulation results show that the reinforcement learning agent achieves task by selecting the optimal one from policies with the proposed method.

In recent years, reinforcement learning has developed rapidly with deep learning and achieves great performance not only in the game playing but also in the continuous control of robots. Reinforcement learning requires exploratory behavior, and action noise is widely used to realize it. Recent researches have tackled exploration problems in deep reinforcement learning by using parameter noise. It has been experimentally shown that parameter noise performs a better exploration than commonly used action noise. However, the methods used so far need long time to update noise distribution or explore uniformly in a huge parameter space by using isotropic noise distribution. This paper proposes a method which improves the update of the noise distribution for faster learning.

In this paper, we introduce a policy search reinforcement learning method with a sparse non-parametric policy model. We formulate policy search as a variational learning problem. A sparse pseudo-input Gaussian processes (SPGP) is placed as a prior distribution of the control policy, then a variational lower bound of the expected reward is derived, which is optimized w.r.t. the hyper parameters and the pseudo-input variables. We conducted numerical simulations and real robot experiments, and confirmed the effectiveness of our proposed method.

A method of autonomous driving is proposed for a disaster search robot with multi-axis vehicle by use of physical modeling and machine learning. The vehicle kinetic motion is realized by physical modeling and the vehicle autonomous driving is controlled by Artificial Neural Network(ANN). ANN is optimized by real number genetic algorithm. The validity of the proposed method is verified by simulation experiments.

Nowadays, autonomous robots are being developed for a variety of situations. However, it is difficult for many of these robots to adapt to environmental changes and unexpected situations. One approach to solving this problem is reinforcement learning. By applying this learning to robot control, a robot can learn and acquire the optimal behavior to achieve a task in a given environment. In this study, we applied reinforcement learning to the task of reaching an upslope goal using our robot modeled on four-legged animals. In simulation experiments, when the robot learned under the condition that it was able to determine the angle and speed of leg-swinging motion as actions, it reached a goal on a five-degree upslope. When the robot learned under the condition that it was able to determine the duty cycle of supporting-leg motion simultaneously with other action variables, it reached a goal on a seven-degree upslope.

In this research, we propose an error recovery system in robotic snap assembly task by learning the result of past assembly tasks. During a snap assembly task, the proposed system can judge the type of error and what adequate motion should be performed for error recovery. In this paper, firstly, we obtain the feature quantities of force and torque data measured during simulated snap assembly tasks. Then, we cluster these data into success and several different failure cases. Furthermore, we try to predict an error in realtime in the middle of a snap assembly task. Finally, we show simulation results of error detection.

In the industrial field, high-speed and high-precision positioning control is required to improve productivity. In order to improve the performance of positioning control, it is necessary to set a control gain suitable for the target machine. As an example of the feedback gain tuning, there is “Iterative Feedback Tuning (IFT)” method, but there are problems such as time for repeated driving. For this reason, we tuned the feedback gain for the mechatronics system using "Fictitious Reference Iterative Tuning (FRIT)". In this method, it is sufficient to drive the target machine only once by performing the offline process. In the experiment, the target gain is decided beforehand, and it is aimed at stable tuning to the target gain by performing five tunings with FRIT. The final experimental result is tuned to a value close to the target gain at all 5 times and shows the usefulness of gain tuning at FRIT for mechatronics system.

It is hard for robots to place foods automatically on a launch box because detecting shapes of foods are difficult. Industrial products are usually standardized; however foods are not done. In previous research, an irregular shaped food detection method is proposed without 3D shape models. However, it needs empirical parameters and thresholds. We propose a new method which uses clustering to graph structure data converted from 3D point cloud data to detect irregular shaped foods. Our method has fewer required parameters than the previous method. We have an experiment of fried chicken serving. The experimental results show 85% success rate.

This research proposes adaptive gripper. A L-shaped nger is used to adjust the initial position error of pegs by pushing twice without any help of vision/force sensors. Besides, an adaptive nger from Festo company is used for adapting the nger for dierent shaped pegs. The ability of the proposed gripper for adjusting the initial pose error(position and orientation) was explored by simulations and physical experiments. The superiority of the proposed gripper was proved by comparing with a two-nger and a three-nger grippers.

Bin picking technology is one of the today's important topics, which realizes efficient production process. In this research, we considered a parts recognition method for small mechanical parts with a complicated shape and occlusions. In this paper, we will report the results of experiments of parts recognition and picking.

We demonstrated picking up the objects using an industrial robotic arm with biomimetic octopus vacuum gripper. The robotic arm controlled press force and moving speed by teaching-playback method. The pressing force is constant by feedback control of robotic force sensor. We measured the acceleration of arm using 3-axis accelerometer and Arduino. The gripper can control the suction and change in flexibility. The volume of gripper decreased when the stiffness increased. The gripper can contact the object, because the robotic arm moved until the constant press force. The robotic arm with gripper can grasp a flat aluminum block, acrylic tubes, and water-filled aluminum laminated foil pouch. These results provide the industrial robotic arm with biomimetic octopus vacuum gripper is useful for the material handling.

We propose a tool-use model which considers relationship between tools and target objects. Robots with tool-use ability are useful in human living space. However, in previous research, robots could not manipulate arbitrary objects with arbitrary tools without requiring any human assistance. In this study, we construct a model that makes robots consider the relationship between tools and objects by themselves and realize robots’ object manipulation with tools. For this purpose, we let a robot experience some tool-use tasks and train sensory-motor data recorded during the experience with deep learning. To let robots consider the relationship, it is necessary to set up the tasks, that include information of four factors: (1) tools, (2) objects, (3) actions and (4) effect. In model evaluation, we analyze whether the robot can detect the relationship between unknown tools and objects.

In order to construct a Brain Machine Interface (BMI) system, it is critical to elucidate a characteristic of brain activities in relation to internal motor representation, namely motor imagery. This study investigates the nature of transfer of motor imagery from one (i.e. left) side to the other (i.e. right) side of the visual field by means of showing enhanced activities of electroencephalographic (EEG) signals. We found that motor-related activities during motor imagery task transferred across visual fields with reversing goggles. Furthermore, the activity persisted after taking off the goggles, demonstrating that visuomotor learning of motor imagery completed.

Cooperation of multiple element motions is important for robots to realize various complicated tasks. Most of the researches focus on realizing a single and complicated element-motion using a motion-generating-model made of deep neural network. In this study, we propose an integration method for those models. We introduce a timing determiner to determine the execution timing of motion, as well as an autoencoder and a recurrent neural network in the model as the novel integration method. We have confirmed that a passing-through-door motion, cooperating multiple element-motions is accomplished by the method.

Recently, a lot of researches on image classification methods based on electroencephalogram (EEG) have been reported with good identification rates, closely 80%. However, most of these results were achieved only in silent environment. If the EEG data for evaluation which are measured in noisy environment were applied to the model constructed using other EEG data which are measured in silent environment, the identification rate would be decreased due to the influence of the environment changes. We have ever developed image classification method which is robust to change in the environment using Wavelet analysis and Deep Learning algorithms. In this research, we recorded EEG data at four different places. We evaluate classification method using recorded EEG data. We discuss about image classification method which is more robust to change in the environment.

The mental rotation task is one of a popular cognitive function program, in which objects, images, and the body are mentally rotated. And, it was reported that training with mental rotation tasks activated a wide range of neural networks of the human brain. Therefore, we have aimed to develop a mental rotation training system for enhancement of cognitive functions and reduction of depression and anxiety. In this paper, we developed a prototype system of mental rotation task training system using tablet devices, and statistically evaluated its function and effectiveness by the reaction time and correct answer rate through the experiments with university students.