Many ankle foot orthoses have been developed to assist the hemiplegic person in walking, but most of them are passive orthoses that can perform the necessary auxiliary functions only in a specific phase during walking. Therefore, we have developed an active ankle foot orthosis that divides the walking cycle into four phases by measuring the ground contact of the heel and the ball of the finger with a pressure sensor, and can set ankle plantar dorsiflexion motion assist depending on the phase.

Parkinson’s disease (PD) typically causes motor and non-motor symptoms (NMS), where Freezing of Gait (FOG) is motor symptoms and Autonomic Nervous System (ANS), including changes in Heart Rate (HR), consider as non-motor manifestations. Most existing walker systems widely focus on FOG analysis without giving much attention to health monitoring feedback during rehabilitation. This study presents a smart robotics walker system (SRW) for HR monitoring and walking rehabilitation. The system framework interacts with a voice-based virtual assistant and simultaneous use of a PPG sensor. The proposed system framework provides an IoT platform by synchronizing real-time heart rates and rehabilitation information of the user. Meantime, PPG data interjects with the efforts of HEBI actuators to stop driving. The system performance presented a strategic analysis of HR changes with a non-PD person and PD patient by 10m steady line of walking. The result showed that HR rapidly increases while performing with PD patients than a non-PD subject, which validates the performance of walking rehabilitation. In addition, the user can know their HR condition in normal and abnormal situations, and the therapist can evaluate the success of the rehabilitation information from the IoT environment. We aim to provide a better rehabilitation IoT platform with SRW considering health conditions like the HR patterns of the user. The system can implement with any existing walker for PD patients.

In this paper, a rocker-bogie mechanism equipped with SEA (series elastic actuator) is used to step oversteps by performing the dynamic motion. High kinetic performance can be expected by storing the kinetic energy of the motor as elastic energy of the spring and releasing it at the appropriate timing. a rocker-bogie mechanism equipped with SEA was fabricated, and stepping over steps utilizing elasticity experimented.

In recent years, with the development of automated driving technology, there have been reports of practical applications of automated delivery robots, which are four-wheeled robots driving on sidewalks. However, delivery robots should not be a cause of traffic congestion.Therefore, it is necessary to have a mobile robot that can run on the roadway at high speed, is small enough not to interfere with other vehicles, is maneuverable, and can reach the front door of the delivery destination.The currently proposed bicycle-type mobile robot has difficulty in maintaining its posture at low speed. On the other hand, the Segway-type mobile robot is extremely difficult to control at high speed.In order to develop a mobile robot that can transition between the bicycle type and the Segway type, a motion plan for a two-wheeled mobile robot with a variable mobile form was developed.

Long distance runners are required to run at a fixed pace in their training. Pacemakers are effective in reducing the aerodynamic resistance and psychological burden of runners, and have greatly contributed to breaking world records and setting good records in middle and long distance races in athletics.Based on the above,it is thought that the performance of runners can be improved by using a pacemaker in long-distance running in athletics, so this study aims to develop a robot that runs at a certain pace to support the training of runners who are trying to run at a specific pace.In this paper, we verified the control method by running the robot on the track at the target speed in the simulation and designed the actual machine.

In this study, we aim to design and realize an attitude stabilization control of a manned electrically-power-assisted bicycle during following a planned path using steering control. In order to realize the purpose, we derive a bicycle model for designing a control system for autonomous driving. In the process, we investigate Fan model, a nonlinear bicycle model. The model is sufficiently precise and catches the bicycle nonlinear dynamics to be used for the stabilizing control of the bicycle equipped with a steer-by-wire mechanism. We will also investigate effective operation support for preventing falls when manned.

A two-wheeled independent drive mobile robot vehicle can be represented as a simple model for nonlinear control, and a trailer model can be considered by connecting these two vehicles. However, it is often treated as a static model, and dynamic inertia problems are rarely considered. In particular, it is important to consider slippage, which can lead to jackknife problems in trailers. In this study, a method of applying the discriminant equation based on the conventional steering wheel model to two-wheeled independent drive mobile robot vehicles based on a non-holonomic system is investigated for slip discrimination.

We are developing a crawler-type robot that automatically inspects outdoor equipment with an autonomous driving function that uses self-position recognition by RTK-GNSS. We have improved the running performance on rough roads with gravel and crushed stones by improving the speed control of the direct drive motor, obstacle detection on the road surface, and self-position recognition configuration.

Automatic inspection by crawler robot requires following path. However, it is difficult to follow the path on road surfaces with large fluctuations in friction and load. We have designed the control of the direct drive motor to achieve path-following running on various road surfaces.

This paper introduces an omni-directional wheel mechanism with active sub-rollers along a Villarceau circle formed in the cross section of a torus/troid. This unique arrangement of the active sub-rollers and the screw-like outline are expected to improve the traveling performance on uneven terrain.

Among mobile robots operated on land, wheel robots occasionally face problems in the running performance over bumps and uneven terrain. In this study, we propose a wheeled robot whose spokes have an independent extension and contraction mechanism using a rack-and-pinion mechanism. The robot is able to move while transforming its diameter using the spokes. The mechanical rigidity, adaptability to terrain, and controlling method of the diameter-variable wheel are verified and discussed by conducting three experiments, which are the lifting of the wheel by the extension of the spokes, the overcoming performance of a step, and the locomotion by the rolling motion using different patterns of spoke extension/contraction.

Virtual pheromone trailing has successfully been demonstrated for navigation of multiple robots to achieve a collective goal using a leader follower scheme. Many previous works use a pheromone deposition scheme which assumes perfect localization of the robot, however there is always some uncertainty due to sensor errors. If the leader robot gets lost, pheromones are deposited at the wrong places. Other robots trailing its pheromones will end up being in entirely wrong areas of the map. We propose a pheromone deposition algorithm which models the uncertainty of the robot’s pose. Experiments in simulated environments confirm that the proposed method models the uncertainty in pheromone deposition for multi-robot navigation.

Camber angle control is indispensable for non-slip running on a curved track in a wheel running system. In this report, we propose a camber angle control trochoidal running system that includes a spherical link that mechanically immobilizes the grounding point as a non-slip omnidirectional running system.

In this study, we developed a crawler block whose gripping force does not depend on the direction of obstacles or the distance between stairs but wraps around the corners of stairs and exerts its gripping force through the deformation of the material, instead of the protruding crawler often used for driving on uneven terrain. the block was fabricated using a 3D printer and its shape was examined. Blocks with different materials and internal structures were fabricated, and their deformation characteristics were determined by load tests. The crawler blocks to be developed are assumed to be applied to the prototype machine made by the previous research [1].

We have been developing motor-tendon-joint series connection type biomimetic actuators to make robots act jumping, running and dancing. To that purpose, we are now developing an actuator with O-ring tensile tendon structure. In the previous paper, we succeeded to apply the energy of landing shock for the next jumping energy in the shallow knee angle area. We also found that the condition of the landing knee angle around 132 degrees is the best angle to use landing shock to the next jumping. Therefore, in this paper, we investigated how to set the landing knee angle to 132 degrees during continuous jumping. We also examined a landing sequence to achieve soft landing with less bounce. The best landing knee angle for the soft landing is 95 degrees.

In this research, we aim to develop a flapping flight robot that can generate large lift force by wings and that realize flight close to real birds. In order to obtain large lift force, it is necessary to give a wing cross-sectional shape (airfoil) to the wing. In this paper, we attempted to manufacture a balloon wing with a shape close to the airfoil using a balloon filled with helium gas. We also developed a flapping flight robot using the manufactured balloon wings and investigated its flight feasibility.

The movement of cerebrospinal fluid (CSF) in the spinal canal shows a reciprocating field formed by heartbeat and respiration. In this study, a prototype CSF phantom device that can generate driving pressures to simulate various pathological conditions and can be used for diagnostic imaging by MRI, assuming different cycles of heartbeat and respiration, will be developed. A CSF phantom with a simple double circular tube structure is connected to two bellows actuators that put a water tank under pressure to simulate the flow of the human body. In the future, time-resolved images taken by MRI can be associated with physical quantities such as pressure and used as training data for AI.

Bacterial chemotaxis occurs by suppressing the directional change and continue straight swimming (run) when approaching the favorable region. The manner of the directional change differs between peritrichous and monotrichous bacteria, which may cause the difference in chemotaxis. The turn angle of peritrichous bacteria tends to be random that change the posture by loosening the flagellar bundle, whereas monotrichous bacteria with one flagellum tend to turn behind by the rotational change of the flagellum. We investigated difference in the chemotactic behavior between peritrichous (Salmonella typhimurium SJW1103) and monotrichous bacteria (Vibrio alginolyticus YM4). The number density of YM4 cells is higher than SJW1103 at the attractant source (L-serine). The numerical simulation where the observed run length and turn angle distribution were taken into account, well reproduced the experimental result. The reason for the higher number density in the monotrichous bacteria is the backward directional changes when receding from the attractant source.

In order to operate a multi-legged robot in muddy area, a special control method for mud walking will be required in addition to a structure that can withstand large loads. In particular, the pressure resistance when pulling up a rod-like leg in a viscous fluid is large, which is a serious problem on mud walking. In this study, we proposed a method of installing a tube that actively supplies compressed air to help the robot and human walk in the mud. In this report, we measured and evaluated the force applied to pulling up in a cylinder and boots moving up and down in a gel liquid.

In this study, we proposed a three-dimensional model which explains the water-repellency function of micro-hairs existing on the surface of the water strider. To derive the three-dimensional model, we used the previous water-repellency models related to the cylindrical and curvature shape of micro-hairs. In the result, the maximum Laplace pressure calculated from the new model was 33.8 kPa. This value was sufficiently larger than the pressure generated on the surface of the middle legs by the rowing motion of the middle legs. This indicates that the water does not enter between micro-hairs, and it means that the surface of the legs is not wet during the propellant motion of water striders.

Vacuum suction is used for gripping objects in factory automation. In the vacuum adhesion, the mechanism becomes large to improve the adhesion force. In this study, we aim to improve the adhesion force per unit area. The Universal Vacuum Gecko (UVGecko) is based on the Universal Vacuum Gripper (UVG). The UVG is a vacuum pad with a deformable skirt. The adhesion forces of the UVG and UVGecko with the same adhesion area were compared. The results showed that the adhesion force of UVGecko was about 11.4 % higher than that of UVG. In addition to the improvement in adhesion force due to gecko adhesion, the deformation of the elastomer membrane was suppressed to prevent air leakage.

A lot of insect-sized flapping robots have been developed for their safety and maneuverability. Especially various designs of flapping robots using piezoelectric actuators have been presented. On the other hand, there are few studies about how the designs of wings affect flapping. In this study, we made the driving system that replicates the flapping of wings using piezoelectric actuators and change the shapes of wings and the driving methods. We measured the amplitude of them with tracking and observe the change in amplitude with changing conditions. By these results, we discussed the impact of the designs of wings on insect flapping biomimetics.

Scientific training is becoming more and more popular in athletics. This training involves analyzing elements of an athlete's movement and feedback the results to athlete. In this study, we focused on load applied on sole of foot. In order to measure the load, it is required that the system does not interfere with running, can measure in three dimensions, and be able to measure anywhere. Therefore, we developed a prototype of a measurement system that satisfies these requirements. As a result of verification experiments, some improvements were found, but it was shown that system could measure load associated with working. Analysis results showed that measured data was valid. In the future, we will improve prototype and establish a new scientific training by this system.

In this paper, we propose a mixed reality device-based system to display golf ball launch information that goes into a hole. The system can show a ball trajectory and animation of the putter head in a real environment to assist the user’s skill acquisition. We utilized our system in the validation experiment to verify whether the visual display of the ball launch information is effective enough in assisting ball control. 12 subjects were divided into two groups: Group 1 with no visual support and Group 2 with visual support. From the results of the experiment, we confirmed that Group 2 was able to hit a ball that stopped closer to the target than Group 1.

In this paper, we propose a system that enables golf putting practice in a mixed reality environment by integrating HoloLens with a vibrotactile haptic device. Using this system, users can hit a virtual golf ball on a virtual green course using a real putter. In the experiment, participants practiced golf putting in a mixed reality environment using this system and we verified whether their ball control skills improved in the real environment. From the results of the experiment, we confirmed that participants were able to hit the ball and enable the ball to stop closer to the target after the practice compared to that before.

In robot motion, collision avoidance is essential to ensure safety, and generation of multiple paths according to the objective is beneficial. Conventional methods have been pointed out to have difficulty in achieving both collision avoidance and optimization, and to have large computational costs. It is also challenging to compensate for safety using deep learning methods. In this paper, we propose a method that can generate an optimal trajectory with obstacles in a short time and guarantee collision avoidance. The optimal trajectory is calculated by updating the latent variables with the collision avoidance posture of the robot obtained as a latent representation of the conditionally generated adversarial network (cGANs). As a result, the optimal trajectory considering collision avoidance could be generated in less time than the conventional method. In addition, We also experimented with collision avoidance using a real robot, UR5e.

In order for robots to perform tasks that humans perform, it is necessary to process multimodal information such as vision and force, just like humans. In this study, we propose modality attention by deep predictive learning that can interpret which modal information is used during the task. A hierarchical model consisting of low-level NNs(Neural Networks) that process each modal information individually and a high-level NN that integrates the modal information is used. Furthermore, by weighting each modal information input to the upper NN with learnable weights and inputting it, the modal information used for motion generation is self-adjustable. We verified the effectiveness of the proposed method in the task of inserting furniture parts that require vision and force. It was confirmed that the modality that attracts attention transitions appropriately, and that stable motion can be generated even if noise occurs in the modality that does not pay attention.

Concentration is closely related to any person, and there is a growing trend to estimate concentration to improve work efficiency and learning efficiency. In order to measure concentration, there is a method of phase analysis using auditory steady-state response (ASSR), but it is time-consuming to calculate. Therefore, the purpose of this study was to measure concentration using power spectrum analysis with the ASSR. The power spectral density of the ASSR was found to increase with the increase of concentration at the F3 electrode. This study suggests that the power spectral density of the ASSR can be a new index for measuring concentration.

In recent years, people tend to be more health-conscious. Due to this, strength training has been attracting attention as an effective way to maintain and improve muscle strength which affects QoL. In this study, we developed a resistance training device for multi kinds of muscle trainings, which is able to set an arbitrary load by motor control. To effectively improve athletic performance with a small risk of injury, the training variables including load, repetitions, and intervals should be determined appropriately based on muscle characteristics such as maximum muscle force. Therefore, we examined the method to estimate 1RM easily based on the load-velocity profile (LVP) to evaluate muscle characteristics such as maximum muscle force. The experimental results show that 1RM (1 Repetition Maximum) can be estimated with high accuracy using several LVP datasets. Finally, we propose a protocol to evaluate muscle characteristics non-invasively based on the experimental results.

Video motion capture needs no markers on subject, and it makes advantage in motion capture of sports. However, video motion capture sometimes fails in estimating some characteristic poses of sports due to the lack of such data in training dataset. In this study, we transfer trained an image pose estimator using an original sports image dataset to use in video motion capture in sports field. We also 3D reconstructed sports motion of taekwondo and artistic gymnastics. We found proposal method is useful when reconstructing motion such as vault motion that is difficult to be reconstructed by some conventional method.

In this paper, we apply the generation and control of whole-body motion based on relative angular acceleration for a humanoid robot, which has been proposed so far, to the generation and control of forward jump, and evaluate the effect of arm swinging movement on the whole-body motion quantitatively from the simulation results. The arm swing movement realized by controlling the spatial acceleration at the tip of the arm in the preliminary motion phase before take-off of the forward jump caused an increase in the flight distance, a change in the angular momentum during the motion, and a decrease in the joint torques during take-off and landing.

For the improvement of Bat Control in the current training, it is difficult for beginners to correct the swing by the information that is the response when hitting the ball. Therefore, VR technology give us quantitative feedback. For the purpose of improving the Bat Control of beginners, we have developed a training system that can obtain visual and quantitative feedback in virtual space. Using the system, we conducted a verification experiment to verify whether it can swing accurately against the target ball by training. As a result of the experiment, there was a tendency to improve the Bat Control by using the system.

In order for robots to move in the real world, they must first be able to understand how to move their own bodies, as there are uncertainties in robot structures, camera parameters, or the tool that it holds. We propose a Dense Image Jacobian Estimator (DIJE) as a method to estimate the image Jacobian online for every pixel in the image. The estimator is based on a Kalman Filter simplified to run on every pixel, and does not need a priori knowledge of the robot kinematics. The proposed method is applied to a markerless visual servoing controller and verified on a musculoskeletal robot.

One of the important advantages of musculoskeletal humanoids is that it is easy to change the muscle arrangement and to increase the number of muscles according to the situation. In this study, we describe an overall system of adaptive body schema learning that can relearn the changes of body schema with muscle addition of musculoskeletal humanoids from a small amount of motion data. We apply our method to a simple 1-DOF tendon-driven robot simulation and the arm of the musculoskeletal humanoid Musashi, and show the effectiveness of muscle tension relaxation by adding muscles for a high-load task.

Transcranial direct current electrical stimulation (tDCS) is a method of non-invasively stimulating the brain by applying a weak current to the scalp. It has been widely used for rehabilitation of paralysis caused by stroke and application to sports training, but there have been few reports targeting the auditory cortex. In this study, we investigated the effects of tDCS on the auditory cortex in terms of both EEG and discrimination ability, using two tasks: pitch discrimination task and oddball task. Although no effect on the event-related potential component P2 was shown in the EEG, the results suggest that tDCS may have an effect on auditory habituation in the pitch discrimination task.

The thinness of a gripper is a big advantage. However existing thin sheet mechanisms have a same disadvantage such as low stiffness. This problem cause reducing its capacity to grasp object. Previously, we proposed active bending sheet mechanism with high stiffness. High rigidity was achieved by deforming the cross-sectional shape as well as the curvature. proposed a mechanism that deforms two thin plates connected by a rotational fulcrum by applying rotational input to them. In this paper, we have realized a bending mechanism with an extended range of motion by controlling the two thin plates independently.

Manipulation of flexible materials in wet conditions is one of the most difficult tasks in daily life and in factory automation. The technique of origami, in which water is applied to a flexible material such as paper or cloth to reduce unwanted wrinkles and create neat folds, is called wet folding. With this in mind, this study presents a robotic hand with wet folding function. For the purpose, the gripper has not only conventional antipodal grasping function but also the functions of pinching by nails, rolling, and water supply. The efficacy of the developed system is validated through several experiences.

We propose a method for direct teaching of a robotic gripper position with zero contact force using proximity sensor feedback. The proximity sensor mounted on the gripper detects a marker position without contact, and the robot track it to generate a continuous trajectory. This system enables us to teach the gripper position in three dimensional positions while keeping the dynamics of the robot and human independent, thus reducing the difficulty of human teaching work and speeding up the teaching process and playback speed. We have confirmed that our method can smoothly teach continuous trajectories in task of drawing a logo. Before applying our method, it took 34.8 seconds to playback the trajectory, but after applying our method, it was reduced to 2.1 seconds.

Currently handling of weak and soft food by robotic system is not widely used in food factories, because it is difficult to pick and place the weak food without any damages by using conventional robotic hands. Some passively deformable soft hand for such weak objects has been proposed. However it is still difficult to apply such weak foods. This study uses a deformable hand with soft bag filled with the air. The air pressure inside the bag is controlled properly. When picking the weak objects quickly using this hand, the objects may be damaged by large acceleration of the picking motion. This study investigates conditions of lowest force to the objects when picking up without any damage. The load force values are measured for tangential direction considering contacting force of the soft bag hand and the weak objects.

In this paper, we propose and realize a spatula mechanism with a folded belt at the tip to scoop and hold a grasped object by bending a thin plate. In addition, a belt tension adjustment function is provided with the same input as that for the stiffness changing of the spatula. This paper discusses the principle of the mechanism, the configuration of the actual machine, and the knowledge of the scooping function obtained from experiments on the actual machine, including its basic operation.

We propose a method for fast estimation of reflection parameters of an object surface using only our optical reflection proximity sensor. The sensor values are the distance to a object surface, the tilt angle, and the reflected light intensity value, and the reflection parameters are searched by comparing them with a ray-tracing simulator values. Specifically, the input values of the simulator are the sensor values and reflection parameters of an object surface, which are optimized by TPE (Tree-structured Parzen Estimator) to minimize the error between the sensor values and the simulator values. This computational procedure has the advantage of being easy to implement since it does not require any other camera devices. However, it has the disadvantage of requiring more than 9 minutes of computation time. To overcome this disadvantage, we replace the entire search process with a suitable NN model. The TPE search takes 589 seconds, but can be accelerated to 3.1 ms by replacing it with the NN model.

Soft robots have a problem that they easily punctured or torn by contact with a sharp edge due to their soft bodies made of weak materials. To solve this fundamental issue, we proposed a robotic blood vessel mechanism that enables the robot to heal their bodies themselves. This mechanism realizes the self-healing function of soft robots like blood closing and healing wounds. We have created some test pieces consists of elastomer and inner blood that reacts by two-liquid mixing. In addition, we constructed a twisted helical type vascular mechanism, which is constructed by the gel produced by two-liquid mixing. In this paper, we demonstrate the concept of the robotic blood vessel mechanism and show the results of the basic experiments.

In recent years, robot technologies for handling a wide variety of products have been attracting attention in factory automation, especially soft robotic grippers. Among them, universal jamming grippers have been the subjects of much researches and developments. However, a unified experimental method has not been established in the past, and it has been difficult to systematize the gripping performance based on the parameters of jamming grippers. In this study, we conducted an analysis of variance to clarify the contribution rate of each parameter to the gripping force.

In this paper, we propose a fluidic variable resistance mechanism as a memory device. Fluid circuits are expected to effectively bring simple intelligence to fluid-driven flexible grippers, which are promising solutions for grasping objects of various shapes; however, a fluidic device corresponding to a memristor has not been realized. Here, we introduced a mechanism in which the resistance changes when the flow rate exceeds a certain amount and acts as a memory device in fluid circuits. The prototype showed a pressure loss-flow rate characteristic with hysteresis, suggesting it can function as memory. As future development, the concept of integrating the proposed fluidic element　into fluid-driven grippers is also described.

Robot hands that can grasp heavy objects are required in various fields such as disaster site and production site. Many researchers developed a wide variety of the high grasping force robot hands. Authors developed robot hands using a transmission mechanism of a ball screw. The feature of the robot hand was high precision positioning and high grasping force. The robot hand had non-linear relation between joint angles and motor angles. This paper proposes a joint angle control method considering characteristic of transmission mechanism of the ball screw. Experiment results show the effectiveness of the proposed control methods using the robot hand.

Drivability is a key aspect of vehicle dynamic performance and comprehensive evaluation is necessary for ensuring drivability quality as such complicated driver operation and vehicle behavior. Furthermore, vehicle control program would be complex for safe and secure vehicle dynamic performance. This paper proposes a novel automated drivability screening system. The proposed system is composed of automated evaluation sub-system and automated exploring sub-system. The automated evaluation sub-system is drivability evaluation by using driver model and PT-VRS equipment to mimic expert driver. The automated exploring sub-system is used to explore feasible region of design space described by control parameters and simulation conditions. To show effectiveness of the proposed system, an example is demonstrated by comparison to expert driver.

Accurate identification of minimum set of dynamics parameters is required for high-precision and high-speed motion control. However, since there is generally some difference between the model used for control system design and the dynamics of the actual target system, the conventional parameter identification method cannot provide an appropriate model for use in control system design. To address this problem, we have proposed a method to obtain parameters suitable for use in control system design by considering stochastic variables. In addition, we have proposed a method to obtain the probability distribution of the parameters and to update them recursively by using the concept of data assimilation. In this study, the effectiveness of our proposed stochastic parameter identification and its sequentialization is verified using a planar 3-link manipulator.

Reinforcement learning (RL) in safety critical domains like autonomous driving requires safe exploration, but conventional safe reinforcement learning methods have low performance at the initial stage of learning. In this paper, we present a RL method that selects action using independent Q-functions on rule-based policy and RL policy to address this issue. In our method, Q-function on rule-based policy is pre-trained using offline data and Q-function on RL policy is initialized randomly. Our method selects action by comparing both Q-functions and increases probability of selecting rule-based action at initial stage of learning. We conduct experiments on driving lane selection task, and find that our approach can improve performance at the initial stage of learning.

Deep reinforcement learning is applied to the pitch control of the UAV equipped with electric ducted fans and thrust vectoring. A controller is trained to maintain the target attitude even when the system model is uncertain. Through the uniaxial rotation experiment, it was shown that the attitude can be controlled by the deep reinforcement learning-based approach with domain randomization.

It is important to obtain the contact force during polishing or grinding processes using a robot to achieve high quality processing. The frequency of the machining sound was correlated with the contact force. In this paper, we propose a method for estimating the contact force based on the sound using a convolutional neural network. Microphones are superior to force sensors in that they reduce the risk of failure and are less expensive. Mel spectrogram was used for frequency analysis of the sound. The results of the experiment showed that the contact force was estimated from the sound. When the contact state was kept constant, the accuracy of the sound-based estimation was better than that of the force sensor-based.