To elucidate the effects of individual finger shape differences on skin sensation, we investigate the effects of finger shape on contact force and area when pressing against a flat plate. The influence of finger shape on the contact force and area when pressing a flat plate is investigated to clarify the influence of individual differences in finger shape on skin sensation. In addition, to eliminate the influence of material properties, a finite element analysis was conducted using an individual finger model with the same material properties. Finite element analysis will be conducted using an individual finger model with the same material properties to eliminate the influence of material properties.

With the increase in the elderly population, the incidence of fall-related injuries among the elderly is rising. To address this issue, it is necessary to estimate the risk of injuries caused by falls. In this study, we developed a simulator that replicates actual falling motions. This allows for the estimation of the impact force on the hands. The simulator is constructed with a 17-link model, and a spring-damper model is implemented in both hands to calculate the impact force. It generates scenarios from the onset of a fall to a few seconds after the body lands. The simulation included scenarios for both young adults and the elderly to measure the maximum impact force on the hands. The results showed that the elderly experience higher maximum impact forces than young adults. This result indicates that the risk of falling increases with age.

In this study, we leveraged Text to Motion (T2M) technology for the facile construction of a dataset aimed at joint angle estimation and verified its validity. Our findings revealed that T2M data could be generated with relative ease, devoid of environmental constraints such as force plates, thereby allowing for the production of variably rich data. However, the current T2M dataset exhibits inferior estimation accuracy when compared to evaluations using AIST models on AIST data and T2M models on T2M data. Discrepancies in minor features between AIST and T2M data, particularly noticeable at endpoints, were identified as potential causes. Addressing these discrepancies by aligning the detailed characteristics of T2M (simulation data) and AIST (real data) closely emerges as a future challenge. We propose that combining datasets or altering learning methodologies, such as Fine-tuning, could be effective strategies for enhancing estimation accuracy by bridging the gap between these data groups.

Low leisure motivation has emerged as a primary challenge for individuals seeking to relieve daily stress. In order to evoke the inclination for rest, we have chosen the widely accepted modality of “music” as a point of entry. However, exploration indicates that compared to “listening to music”, actively engaging in “playing/composing music” is more effective in stimulating leisure motivation and facilitating superior relaxation. To address this, we have developed a music-generation wearable device tailored for use by amateurs. By activating users’ self-efficacy, the transition from passive listeners to active performers is facilitated, thereby enhancing daily leisure motivation. Through experiments, we have preliminarily validated the high system usability of the prototype and confirmed its effectiveness in continually improving musical self-efficacy.

This research aims to develop a learning support robot with a 2-DOF laser pointer and a rotating platform, connected to a wireless network and operated remotely. In the experiment, we calibrated the pointing device to improve the pointing accuracy at the target point, which was improved by 34%. The average error of pointing accuracy was within 1.5 [mm], and the maximum error was 3 [mm] in the integration experiment, which was designed to be used by students and teachers.

In this study, we propose a method of imitation learning that employs a Transformer to capture the transitions of multiple focal body parts in motion teaching between human and robot with Embodiment gap, such as weight balance, joint degrees of freedom, and range of motion, between humans and robots. This approach aims to characterize the actions in the context of motion teaching, where the Embodiment gap of the agents exist. By focusing on the transitions of specific body parts, our method seeks to effectively learn and replicate complex movements.

In public spaces, there is a need for information transmission methods that increase self-involvement which treats external information as own business. We focused on the direction of information transmission and developed a eye-projection calling robot to increase self-involvement. This robot can enhance a sense of information transmission through directional speakers and gaze expressions. However, it is not designed with a sense of distance from the target user in this robot. In this study, we focused on the mechanism of human focus adjustment and proposed an attention expression method that generates the pupillary distance to the target user based on viewing distance and binocular disparity.

In online communication such as meetings and classes, many participants can connect anywhere,and any number of people can participate. Although online communication is highly convenient, it is difficult to share a sense of unity and interactive atmosphere because the monitor is perceived as a spatial boundary. In face-to-face communication, nonverbal behaviors such as facial expressions and gestures play an important role in sharing a sense of unity and interactive atmosphere. To improve a sense of unity in remote communication, many avatar robots that represent nonverbal behaviors of remote user have been developed. However, there is no approach in creating interactive atmosphere such as activated communication through nonverbal behavior using multiple avatar robots, because the previous research has assumed a scene where the user and the avatar robot faced each other one-on-one. In this study, for the basic research in realizing smooth communication between human and avatar robots, we focused on the number of participants in remote communication and analyzed the degree of activated communication during free conversation. The results demonstrated that group of 3 was most activated in online communication.

In this study, we perform form evaluation for weight training based on skeletal point information obtained from a camera. Performing weight training in the correct form is important for improving muscle strength and reducing the risk of injury. However, weight training is usually performed without an instructor. Therefore, we propose a system that can evaluate trainee's own weight training form using a camera. In existing systems, the positional relationship between the camera and the person is fixed. On the contrary, our system acquires 3D skeleton point information from the video, and thus the system is robust to variations in the positional relationship between the camera and the person.

Cerebral organoids, that is cultured from iPS cells, are known as a counterpart of the human brain. Interestingly, some researchers reported cerebral organoids exhibit adaptive reservoir computation ability with biological neural networks. Therefore, we expect that the cerebral organoids can be used as a processing unit for robots, and we can develop novel intelligent systems using cerebral organoids. In this study, we aim to develop a cartridge-type measurement system so that we can measure the neural activity of a cerebral organoid and observe with a microscope simultaneously. As a result, we confirmed that we got time developments of neuronal firing and computation results of spike sorting. We observed a cerebral organoid inside the chamber with a microscope at the same time of activity measurement. In future work, we expect we can implement cerebral organoids into the robotic systems via the proposed cartridge-type measurement system.

This paper describes a lower-limb exoskeletal actuator-less mechanism that we are developing to assist the standing movements of rehabilitation patients. First, the three main actions of assisting standing movements are described, and then, the buttock release mechanism is reported, including an explanation of the principle of the mechanism using a simplified model and the measurement results of the buttock release force using sensors.

Various lifting motions are expected at work sites, and lifting heavy luggage may cause lower back pain. Work-related back pain is a severe problem worldwide. In addition, the number of lower back pain is increasing. Therefore, Authors developed the lower back assistive device with pneumatic assistive system to reduce the rate of lower back pain cause. This device is lightweight to reduce the burden during lifting motions because it does not have external power supply such as battery or air tank. In this paper, we have developed a lower back assistive device that is suited to the actual conditions of the workplace by generating assistive force in accordance with human movements, adapting the device to various body sizes, and reducing the mass of the device to enhance the assistance effect.

Mirror therapy using mirrors has been used to restore upper limb function in stroke victims with hemiplegia. We used brain measurements to confirm whether the inverted prism can produce brain activity similar to that of the conventional mirror therapy technique. The results showed that both inverted prism and mirror therapy tended to enhance the inhibition of μ-rhythms. We also found that the ERD of the unaffected hemisphere was reduced in the low beta frequency band, suggesting that the inverted prism may induce brain activity similar to that of mirror therapy.

Many studies have shown that complex tasks such as walking are controlled by groups of coactive muscles or joints, or synergies. In the analysis of muscle synergies, it is important to reveal their biomechanical functions for specific tasks. However, how the function of muscle synergies is altered by walking speed has not been studied in detail. In this study, we analyzed the relationship between muscle synergies and kinematic synergies when walking at different speeds. In addition, we extracted the kinematic synergies from the angular velocity data in order to make it easier to correlate the muscle synergies and the kinematic synergies.

Hip Exoskeleton assist robots are capable of applying strong assistance force to users. However, they often lack back-drivability, leading to the generation of excessive resistance force on the person. Therefore, while there is research aimed at enhancing back-drivability in the direction of leg lifting and lowering during assistance, studies incorporating back-drivability for movements in the leg’s abduction and adduction direction, as well as the internal and external rotation direction of the foot, are limited. In response, we have implemented mechanisms with adjustable compliance for assistance directions and mechanisms that allow for free movement in permitted movement directions. By doing so, we designed mechanisms that possess back-drivability in each direction, and this paper elucidates their utility.

This study proposes a novel reach extender using Nankin Tamasudare, a traditional street performance tool. The device can support elderly people and pregnant women who have difficulty standing up. It can quickly and intuitively reach the maximum distance of 3 m; allow users to pick up objects without standing up. The strong point of this mechanism is the ease of design. The segments of the telescopic mechanism are comprised wood and aluminum. It weighs only 393 g including the gripper, which is operated by a small DC motor that works as a trigger. This gripper acts two roles: catch and release depending on the position in it. Experiments confirm its quick and smooth telescopic motions, and easy manipulation. Moreover, we demonstrate the ability to catch an object in a high place; to slide into gaps and catch an object.

Humans have various abilities. However, we are unaware of the existence of abilities that are not needed in our daily lives because we do not have the opportunity to use them. Conversely, by setting up a situation where people work under conditions different from their daily lives, it is thought that we can discover new abilities that are hidden within us. The authors will refer to this field of research as “Ability Mining.” Research on ability mining is thought to be beneficial in the long run because it leads to understanding “what kind of abilities people have.” This paper proposes a research concept of ability mining. As an example, this paper introduces a study that focuses on the ability of operation by toes, and reports on the results of subjective evaluation.

The Stretch-Shortening Cycle exercise is an exercise in which the muscle is fully powered by performing tensile muscle activity before shortening muscle activity. This exercise can be used by performing the movement in the opposite direction of the main movement before the main movement. In the case of jumping, it can be used by performing a vigorous sinking motion before the main motion of jumping straight up. In this study, we investigated whether the incorporation of the SSC exercise in jumping and badminton swinging, in which the SSC exercise can be best embodied, improves athletic performance, and aimed to create a system.

We aim to elucidate canine super-sensing, which is the ability to notice changes in human emotion. If we can elucidate canine super-sensing, we believe that it can be applied to various fields as a new sensing principle. For the purpose of elucidation, we have fabricated a device that presents stimuli to multiple canine senses. However, with the previously proposed method, the canine could not concentrate on the presented information due to the discomfort of the visual stimuli, and the canine did not concentrate on the presented stimuli for 30 seconds. In this paper, we describe how we improved the method of presenting visual information and automated the apparatus to shorten the experimental time.

The authors are developing a training method to provide rewards from a canine-mounted feeder. Rewards from a suit-mounted feeder that does not limit the canine’s range of behavior to the vicinity of the machine or person performing the feeding are essential for the long-range behavior guidance of canines using light stimulation that the authors’ research group is working on. However, there is no suitmounted canine feeding device with optical stimulation, and it was unclear whether such a suit could be used for canine training. In this paper, we developed a canine suit equipped with light stimulation and a feeder. Using the developed suit, we showed that it is possible to train canines to guide their behavior using light. We also showed that the developed suit could guide the canine to the desired location, although there were some areas for improvement.

A surface contact model was constructed to calculate the optimum grasping position and force. A total of four surface contact models have been studied so far. In this study, a boundary non-uniform distribution model and a total non-uniform distribution model are added, and the minimum gripping force is examined based on a total of six models. The theoretical values calculated based on the models are compared with the measured values using the robot hand. As a result, although linearity was confirmed for both the theoretical and measured values, it was found that a difference existed between the theoretical and measured values.

Most of the processes in the sericulture industry strongly rely on manual labor. In order to solve these problems, we propose to replace the current manual processes with robot automation. In this study, we propose a soft gripper that can pick and place silkworm larvae without damaging them. Gripping force measurements were performed for a rising gripper driven by a silicone belt. We measured the gripping force for various materials and dimensions of the silicone belt. In the gripping experiment, the objects to be gripped were a model and a silkworm, and six types of silicone belts were used.

In order to optimally grasp a hamburger steak with a three-finger robot hand, we simulate how the grasped hamburger steak is deformed. Young's modulus of the hamburger steak was examined using a digital force gauge, and the deformation of the hamburger steak was simulated using the examined values. In the simulation, the position of the fingers of the hand was moved in order to compare the difference in deformation depending on the grasping position. The method of measuring Young's modulus of hamburger steak using a digital force gauge is not limited to hamburger steak, so it can be used to measure Young's modulus of other flexible foods.

In this paper, we proposed a robotic hand equipped with self-locking mechanism for handling heavy agricultural products such as radishes in order to improve the situation of labor shortage at fruit-sorting facilities due to the declining birthrate and aging population. The proposed robotic hand does not require active actuation,such as motor or air compressor，by taking the advantage of the locking mechanism．The locking mechanism is incorporated into the joints of the hand，and the hand opens and closes by contact with the environment We conducted experiments on handling heavy vegetables such as radishes and other heavy objects,and compared the performances of the hand with and without the restraint mechanism．

Wire-driven robots have many advantages, such as weight reduction of the manipulator. However, the further the driving part goes toward the tip of the manipulator, the more complicated the wire paths become, which complicate the calculation of the wire length.

To solve this problem, we have developed a mechanism in which all joints are independent from each other, which makes it easier to calculate wire lengths. We developed a three-jointed robotic finger using this mechanism. Then, we evaluated the motion and torque of this. As a result, We confirmed the independence of the joints. However, errors occurred in the torque at each joint. We consider that this is due to differences in the stability of the motor torque and in the length and path of the wires at each joint. Finally, three robot fingers were combined to form a robot hand to confirm the usefulness of this finger.

Jamming grippers utilize the jamming transition phenomenon and can grasp objects by simply adjusting the negative pressure, regardless of the object's shape or posture. However, in order to place the object to be grasped in a given position and posture, a gripper, sensing, and jig for re-holding are required accordingly. A gripper that can manipulate the grasped object in the hand would solve the aforementioned problems. In this study, we propose a two-layer structure of a universal gripper to realize in-hand manipulation with a universal gripper. We also attempt to control the position and posture of the grasped object by controlling the pressure distribution using the two-layer structure of the universal gripper. This paper outlines the two-layer structure of the universal gripper, presents the concept of in-hand manipulation by pressure distribution control, and reports on the experimental situation with the developed gripper.

This paper describes a vacuum suction pad with a tactile sensor capable of evaluating and predicting the fall of a grasped object. First, it provides an overview of the research background, the utility of the suction pad, and related studies. Next, it focuses on the principles of the conductive sponge substrate and amplifier substrate, which constitute the developed device, as well as the tactile suction pad fabricated by integrating them, highlighting improvements made since the previous year. Following that, experiments are conducted to assess whether the improved device can accurately measure ”the grasping state of an object,” and the results and discussions of these experiments are presented. Finally, the paper discusses trajectory planning based on the adhesion information obtained from the experiments.

This paper proposes a 3-finger robot hand with an active ball at each fingertip, which is capable of bin picking and in-hand manipulation. The active ball is a rubber-coated steel ball, which is rotated by two motors located orthogonally. Appropriate rotations of the three balls enable a grasped object to rotate in all directions. A parallel link mechanism is adopted for picking objects while avoiding contact the fingers with surrounding objects; a vacuum type is adopted for picking up the objects; image processing is used for detecting the feature point. We demonstrate the apple shipping process to verify the feasibility. Experimental results show that the hand was capable of picking an apple even if it was located at the corner and that the active balls rotated the apple to orient the stem upwards using image processing.

This study aims to realize a pick-and-toss operation that can replace pick-and-place, thereby allowing the robot to expand its work range and perform tasks more efficiently. While pick-and-toss improves the speed of object arrangement tasks, the placement environment influences the success or failure of the tossing operation. Therefore, to achieve quick and accurate object arrangement tasks, we propose to determine the preferred operation from pick-and-place and pick-and-toss tasks by considering the task difficulty estimated from the target placement environment. The proposed method can simultaneously acquire the tossing trajectory through self-supervised learning and the task decision policy through a blue force search. Our experimental results demonstrated the effectiveness of the proposed method through simulations and real-world experiments of arrangement tasks for several different rectangular shapes.

This paper proposes a variable stiffness spatula mechanism that can be deformed into a U-shape. This mechanism is a spatula mechanism that scoops up and lifts the object by increasing the cross-sectional quadratic moment of the thin plate. The principle of this mechanism, the configuration and basic operation of the prototype, and the stiffness measurement experiment using a prototype spatula and its results are described.

The tasks carried out by robots instead of humans are becoming increasingly challenging. To meet these demands, robots need tree-type manipulators which have high h degrees of freedom and branches like humans. Control theory using products of exponential (POE) and chain matrix enable control of such manipulators. However, no control of a tree-type manipulator has ever been performed with POE and chain matrix. Therefore, in this study, we control tree-type manipulator with POE and chain matrix. We compared the target and measured values of the tip position for evaluation.

As service robots play a more active role, humanoids are expected to play a more active role. The advancement of humanoids requires the development of lightweight, flexible, and redundant robot arms. In this study, we explore the integration of tensegrity structures to enhance mechanical flexibility in robot arms for humanoids. Our focus lies in designing a lightweight, flexible, ultra-redundant robot arm that incorporates tensegrity module structures. Our goal is to control the behavior of the robot arm by applying the Product of Exponentials theory. As the first step, in this paper we developed the wire-driven tensegrity module and assessed its range of motion through practical measurements. As a result, the developed wire-driven tensegrity module has about twice the range of motion of an air cylinder-driven module.

Preemptive adaptation by using proximity sensing has been developed for rhe robots in unknown environments. We forcus on contact transition motions with impact reduction for robot arm and finger which have insufficient back drivability to be executed force control. Comparing simulation and experimental results, we examine the differences in the responsiveness and strokes of the two manipulators, and the motions are evaluated in terms of reduction of contact force and relative velocity. We also propose two control methods to generate same motions for manipulator consisting of a finger and arm in serial construction which have different characteristics. In addition, we implement the methods for the manipulator and evaluate the motions.

In this research, we developed a telescoping arm using a single cylindrical spring that can be rolled up and becomes a cylinder when extended. It is a simple mechanism that only expands and contracts. An iris robot hand is mounted on the tip of the arm. Additionally, in order to reduce the weight of the tip of the arm, a three-fingered iris robot hand is used, a motor is placed at the base of the arm, and a wire is used to control the opening and closing of the iris robot hand.

In conventional image-based classification methods for micro screws, screws in trays are identified one by one, which is time-consuming and labor-intensive. In this study, we succeeded in detecting and classifying multiple screws simultaneously and with high accuracy by learning with the object detection algorithm YOLO. When screws overlap, the focus of the image is blurred, but the blurred image is also learned to improve the recognition rate.

Due to labor shortages, the number of service robots being installed is steadily increasing. The main functions of service robots include mobility and grasping. However, doors installed between rooms and in storage spaces become obstacles, limiting the environments where service robots can be introduced. Therefore, to relax this constraint, it is necessary to equip service robots with the functionality to perform opening operations. The operation of opening doors with manipulators has traditionally focused on swing and sliding doors whose handle’ trajectories are relatively simple when the doors open. This study proposes a Model Predictive Reinforcement Learning (MPRL) for the operation of opening folding doors, which have not been conventionally addressed.

We propose a fast path planning method for a robot arm using quantum annealing. Path planning is one of the key issues in robotics, and with the increasing demand for service robots in recent years, fast path planning methods in dynamic environments are strongly required. In this paper, we submit the path planning problem into a quantum annealing-type quantum computer by converting the path planning problem into an integer optimization problem and finally to a quadratic unconstrained binary optimization problem (QUBO), with the aim of reducing the computation time. Numerical simulation shows that one path can be obtained for a 3-DOF robot arm in an average time of 80 [μs].

In the context of a global labor shortage, the installation of robot handling systems has been steadily increasing. However, these systems are constrained by limitations in three-dimensional measurement capabilities, particularly the challenge of accurately measuring specular and transparent objects, which restricts the range of objects that can be handled. In this study, we focus on the capabilities of a hand-eye robot system that can capture images of a scene from various poses during object transportation. We propose a method that enables the three-dimensional measurement, grasping, and transportation of specular and transparent objects without significantly increasing the execution time of handling tasks. The proposed method involves performing semantic segmentation on images captured from various poses, followed by the application of shape from silhouette to conduct three-dimensional measurements based on the segmented images.

In this study, we focus on the robotic swarm systems which consist of decentralized mobile robots only having local communication ability. By implementing coupled oscillator system to the individual robots and introducing wave like interaction, robots can estimate global network structure information only using local observation. Under wave type interaction, new edge between robots increase energy of whole system. We derive mass-spring model of this phenomena and propose method to prevent energy increasing.

We propose an autonomous movement control method that aims to maintain sensor coverage of the entire multi-robot system as much as possible, even when some robots have sensor failures. To reduce the degradation of the sensor coverage, the failed robot's position is moved to the position of the robot that does not contribute to the sensor coverage. Computer simulations confirm the effectiveness of the proposed method.

In this study, we aim to develop a method for maintaining the cooperative behavior of multi-robot systems even when the external sensor of a robot fails. We develop a method for the failed robot to estimate the relative position of the neighbor robots. Assuming that only the relative positions between robots can be observed by external sensors, we estimate the relative orientation of robots based on the displacement of the failed robot. The main causes of estimation errors are discussed through computer simulations.

A system is expected to facilitate collaboration among multiple construction machines or robots, enabling them to adaptively perform various tasks in disaster sites and unknown environments. This paper proposes a flow that enables adaptive coordination in unexpected scenarios, based on an understanding of the environment and the functional expressions of robots, utilizing GPT-4 and GPT-4V. Through experimentation, we will confirm that the proposed flow are able to adapt to an unforeseen environment, particularly path obstruction via robot experiments. Furthermore, we investigate the validity of the proposed flow by evaluating the experimental results that change the robots’ functional expression.

In recent years, the need for multi-robot systems in an agricultural field has been increasing due to various factors including labor shortages. One of the important technics for applying multi-robot systems to agricultural works is a path planning that generates suitable paths for agricultural field for each robot. In this study, a path planning algorithm using both graph representation of the field based on a graph theory and path generation by Ant Colony Optimization is proposed for equalizing and minimizing the working distance of each robot. Experimental simulation results confirmed that the paths can be generated with a small error in the working distance for each robot, taking into account the work on all ridges and movement between fields.

The paper addresses a social cooperation algorithm for swarm systems through scramble crossing simulation. In order to facilitate scramble crossing simulation, an appropriate balance between egoistic and altruistic behaviors is required. In this study, weight parameters for egoistic and altruistic behaviors are represented as individual strategies corresponding to personality. The proposed social coordination algorithm is based on imitating the strategies of other agents that show high performance while including appropriate noise. Finally, the effectiveness of the social cooperation algorithm in scramble crossing simulation is verified by flow rate as the global performance for varying agent density.

This study proposes a social network-based approach for quantitative analysis of a robotic swarm. Social network computes the relationship among a group of individuals using a complex network methodology. This method shows the temporal dynamics of the number of communities forming a swarm. In addition, this paper discusses collective behavior with respect to the statistical variance in the directions of robots. These metrics were applied to the path formation behavior of a robotic swarm. The results showed that the robots do not change their number of communities significantly while achieving path formation. Furthermore, the robots belonging to the same communities tend to show velocity alignment. These behavior features implied the correlation to the task performance of a robotic swarm.

In this study, we designed a navigation control system for swarm robots using a synchronization pattern of coupled van der Pol equations. The designed control input consists of three forces, which are cohesion, alignment, and separation. The cohesive force is constructed using the synchronization patterns of the van der Pol equations. The method can control translational and rotational motions of swarm robots by changing parameters of the control law. The effectiveness of the control law is confirmed by real swarm robots.

Grid-maps are often used by autonomous robots to navigate in environment. However, grid-maps have a limitation that they only contain information about free space and obstacles in the map. They do not contain semantic information about the obstacles and the place. Semantic SLAM has been proposed to solve this problem. However, sometimes, replacing the existing grid-map by a semantic map is not feasible as it requires changing the underlying localization and navigation module. To this end, in this paper, a method to embed semantic information in existing grid-map is proposed. The algorithm can use a monocular camera and embed semantic information in the grid-map. The results are evaluated in a simulation based environment.

Robot navigation is a critical component of mobile robots. Robots are typically controlled autonomously, or using direct commands through keyboard, joystick, or mouse. Traditional path planners like RRT (Rapidly Exploring Random Tree) algorithm find it difficult to generate a path though narrow passages in the map. In this paper, we improve the traditional algorithm which takes a long time to find a path through the narrow passages. Our algorithm can speedup search though narrow passages. The results are evaluated in a simulation based environment.

An autonomous mobile robot system is developed with a main sensor of a LiDAR and a PC. Measurement and control programs are described in ROS. For the first test in room environment, the robot autonomously drives along an auto-planed route on the map generated by SLAM. In a case of detection of interrupting objects, the operating programs re-generate an alternative route and the robot avoids the collision. The same movement programs are executed for a drive on an outdoor short passage in daytime. Autonomous movement is not achieved by using an indoor-use LiDAR but by a light resistant LiDAR. The suitable LiDAR should be selected for this development.

Localization is important for autonomous mobile robots. Since localization is performed by matching observation data with a map, it is necessary to have a map which appropriately reflects the environment. Environmental changes cause differences between the environment and the map, which break down localization. In this paper, we propose localization using multiple map for robust localization against environmental changes. In this experiment, we prepare four conditions that have a negative impact on the localization of the stereo camera. We build multiple maps on all conditions. We verify that the use of multiple maps may contribute to robust localization against environmental changes by comparing localization using single map and localization using multiple maps.