In the rehabilitation of hemiplegic patients, quantitative training by the patients themselves is lacking. Rehabilitation using robot technology has been widely used, but quantitative training by the patients themselves is still lacking because it requires a physical therapist to operate and watch over them. To support quantitative training, we have developed a gait trainer that can predict future walking speeds and reproduce increases and decreases in walking speed in real time by controlling a motor, and can follow the walker. The increase or decrease in walking speed is reproduced. In order to implement the real-time prediction model we are developing, we control the walking trainer and reproduce the increase and decrease of walking speed.

Autonomous mobile robots operating indoors need to use elevators to move between several floors. However, it is not easy for the robot itself to accurately recognize the current floor. Therefore, this study proposes a system that uses a barometric sensor to estimate the robot's floor position and moves autonomously using an elevator. We confirm that the proposed method enables the robot to accurately identify the floor and move autonomously by real experimental result. In addition, the system was able to accurately identify the floor even after a certain period of time by introducing a system that accurately identifies the floor even if the air pressure changes with time.

As global resource demand and environmental issues escalate, transitioning to a circular economy is essential. This study, part of a NEDO project, develops an autonomous sorting system for electronic waste in home appliance scrap yards. Traditional localization methods like SLAM and GNSS struggle in dynamic environments. To address this, we propose an artificial landmark-based approach. The system uses vertically stacked geometric landmarks, enabling robust detection and self-localization within a 30 m range. Experiments confirm that localization accuracy improves when landmarks are spaced further apart. Future work will focus on reducing detection errors and stabilizing performance in scrap yards. Additionally, we will implement a switching mechanism that uses either camera images alone or a combination of LiDAR and images based on landmark distance.

Multiple service robots should serve efficiently with each other in large-scale facilities such as shopping malls. The task allocation becomes difficult when the timing and place of appearance of the guests and their destinations are unknown in advance. In our previous work, we proposed a generic algorithm (GA) base task allocation method with a topological map. However, it is hard to develop this method to optimize both the task allocation and behaviors of the service robots. In this paper, we present a deep reinforcement learning based task allocation method with Deep Deterministic Policy Gradient (DDPG) instead of GA. We demonstrated that the DDPG based task allocation can be performed in a situation of six robots and six guests.

The performance of legged robots is dependent on installed leg mechanisms. Among various leg mechanisms, one-degree-of-freedom link leg mechanisms have benefits of light weight and ease of control. However, the design of the link leg mechanisms is tough because analytical design of the link mechanisms is difficult due to their nonlinearity. To address this issue, we have developed a novel design method for one-degree-of-freedom link leg mechanisms using a genetic algorithm. The effectiveness of this method was verified in our previous research using single-objective optimization. In this study, we have extended it to multi-objective optimization. The link leg mechanism designed through the proposed method was fabricated, and it demonstrated the intended motion and performance properly.

In recent years, the number of farm households in Japan has been decreasing due to the aging of farmers and the shift away from agriculture by young people, and automation and labor saving are required as a solution. In this study, we propose an automatic harvesting robot for the purpose of automating harvesting, one of the tasks in Citrus unshiu cultivation. In this paper, we report the results and discussion of harvesting experiments conducted in a simulated environment using an end-effector. We also report on a crawler-type mobile cart that can be mounted on the robot and automatically travel on slopes and uneven terrain in the field.

A method is being developed to improve the identification accuracy of side-scan sonar. By performing a Fourier transform on side-scan sonar data and comparing the results, differences in frequency components can be observed. To improve the identification accuracy, it is necessary to understand the ship’s movement and apply corrections before performing the Fourier transform. The objective of this research is to correct side-scan sonar data using information from RTK positioning. The corrections target three of the ship's six degrees of freedom: roll, pitch, and heave. Experiments were conducted in two locations: at sea and on land. To derive an appropriate correction formula, accurate information is required not only for pitch but also for roll and heave.

This study examines methods for efficiently collecting environmental information, such as obstacle size and location, as an initial response to disaster occurrences. Specifically, we develop and evaluate a decentralized exploration algorithm using multiple mobile robots in an unknown environment. With disaster sites and hazardous areas in mind, we propose a method to explore the entire environment quickly and effectively. In this method, homogeneous autonomous distributed robots determine their exploration direction by combining two clustering techniques based on their self-location and exploration status. Additionally, when robots are in close proximity, they perform local communication to share information. Furthermore, after the exploration is completed, clustering and ellipse fitting are applied to the observed obstacle point cloud data, enabling the estimation of environmental information even from incomplete obstacle data. To verify the effectiveness of these methods, we conducted simulations.