Results of the FY2025 MITOU IT Program
Release Date:Sep 30, 2026
In-page Links
Results of the FY2025 MITOU IT Program
PM: IGARASHI Yuki
A Subtitle-Based Foreign Language Learning Application for Anime and Drama Content
In recent years, an increasing number of learners have been exposed to foreign languages through video content. In particular, younger generations often become interested in language learning through the movies, anime, dramas, and online videos they watch in their daily lives rather than through traditional school materials. However, in many cases, learning ends with simply watching the content, and even when unfamiliar words are looked up, they are often not reviewed continuously or retained effectively.
This project developed “ViewLLA,” a foreign language learning application that transforms video viewing from passive entertainment into an active learning cycle of “watching, saving, reviewing, and speaking.” While watching subtitled content such as anime and dramas, users can tap on words or lines in the subtitles to save them. The saved content is automatically reflected in vocabulary lists and quiz functions, allowing learners to review through cloze exercises and speaking practice. In addition, the system can generate subtitles for videos without subtitles using speech recognition technology, enabling them to be used for language learning as well.
Furthermore, the system utilizes translation APIs and speech recognition AI to support subtitle generation and learning assistance. Rather than treating videos as something to consume passively, the project aims to naturally connect video viewing with active learning behaviors and create a new language learning experience that is easier to continue in everyday life.
Creators: JIANG Lijia
Development of a Programming Education Support System that Dynamically Generates Teaching Materials Based on Students’ Interests
In this project, we developed an educational support system that automatically generates and provides learning materials and projects based on each student's everyday interests and preferences. The system enables students to engage proactively in learning while connecting their personal interests with programming.
Creator: SHINAGAWA Asahi, SHIBATA Toma
Development of a Musculoskeletal Analysis System for Injury Prevention in Sports
In this project, we developed a system that integrates camera-based human pose estimation with various sensor data, including GPS, accelerometers, and insole sensors, to perform musculoskeletal analysis and visualize the loads on joints and muscles. The system aims to provide a foundation for supporting injury prevention and performance enhancement in sports, while also enabling future applications in medical and rehabilitation.
Creator: FUKUSHIMA Ryuto, KAJIO Naoya
PM: INAMI Masahiko
Development of an Agent-Based Facilitation System for Structuring Discussions
In this project, I developed "Recogra," a real-time discussion visualization system that enhances facilitators’ situational awareness during group discussions. Recogra is a web application designed for face-to-face group discussions that transcribes and structures audio recorded in real time from devices placed on each group’s table and displays the results on a facilitator’s tablet.
In face-to-face group discussions, facilitators need to circulate among multiple groups while monitoring the progress of each discussion. However, as the number of groups increases, the time available for each group becomes more limited, making it difficult to understand the context when joining a discussion midway. In addition, during whole-group sharing sessions, information is often selectively conveyed through representatives’ reports, which can lead to the loss of minority opinions and discussion processes.
To address these challenges, I implemented a timeline view that displays discussions chronologically and an issue map that provides an overview of discussion structures to support facilitators while they circulate among groups. I also developed an illustrated summary feature inspired by graphic recording to support whole-group sharing.
Creators: TANAKA Kai
A Modular Social Robot Platform for Self-Expression
In this project, we developed a platform for a modular social robot, Katakko, which users can freely configure and assemble according to their preferences and intuitively control through their own body movements. To address the limitation of conventional ready-made robots, where reflecting individual preferences is difficult, we enable morphology design and motion control without requiring specialized knowledge. The system consists of (1) hardware based on block-shaped modules with a claw-type connection mechanism, (2) a motion translation method that performs camera-based body motion recognition, robot structure estimation from images, and automatic mapping to robot joints, and (3) an integrated software platform using a web interface and a Unity environment. All components are designed to be built from off-the-shelf and 3D-printable parts, with open-source deployment in mind. By integrating the full loop of assembly, configuration, operation, and reassembly, the system enables intuitive trial-and-error in designing original robots. Demonstrations at Maker Faire Tokyo 2025 and The Lab. showed that a wide range of users, including children, could enjoy both designing robot forms and controlling them with their bodies, indicating that the system provides an effective foundation for using robots as a medium for self-expression.
Creator: YOKOI Sotaro, SHITANDA Naoki, MATSUO Kento
An XR Application to Make Practicing Guitar Alone More Enjoyable
SoRock is an XR application that enables beginner guitarists to enjoy and continue practicing on their own. While the guitar is an appealing instrument that allows for self-expression and communication with others, beginners often face several barriers: the songs they want to play are too difficult, practice songs in learning materials do not match their preferences, three-dimensional finger placement is difficult to understand from flat instructional materials, and practicing alone can feel isolating.
SoRock addresses these issues by using source separation and chord analysis to automatically generate practice scores from users’ favorite songs, while also allowing the difficulty level to be adjusted according to their skill level. In addition, virtual hands and picking guides are displayed in XR space, enabling users to understand chord fingerings and strumming motions three-dimensionally. Furthermore, by combining ensemble performance with virtual band members, live-stage effects, and scoring, SoRock provides an experience that helps users recognize their progress and stay motivated.
User evaluations showed that participants repeatedly practiced by being motivated to perform their favorite songs and challenge themselves through rankings, while also gaining a sense of improvement and continuing practice voluntarily. SoRock transforms guitar practice from monotonous repetition into a highly entertaining learning experience in which users can play their favorite songs, perform with a band, and enjoy their own growth.
Creator: TANAKA Sora
PM: OKA Mizuki
Automated Construction of an Urban Infrastructure Database for flood forecasting Using Machine Learning
In this project, I aimed to automate the process of creating an urban infrastructure database, which serves as the computational foundation for an inland flood forecasting system, by using machine learning. flood forecasting requires infrastructure information on roads, sewer systems, and rivers to be organized into a database; however, this process has traditionally relied heavily on manual work in GIS and took about eight months even for central Tokyo alone. By combining GIS software (ArcGIS) with machine learning extensions, I significantly improved the efficiency of this process.
Creators: TAKEUCHI Nozomu
Creating an XR Rehabilitation Space Where Everyone Can “Have Fun Together”
“If only I had done more rehabilitation back then…”
This project aimed to address the loneliness and boredom caused by the repetitive, monotonous movements often involved in rehabilitation by using XR technology to create a rehabilitation space where patients, family members, therapists, and others feel motivated to participate and enjoy rehabilitation together.
I implemented a system in a standalone application that runs directly on a head-mounted display, enabling multiple users to communicate with each other over a local area network without requiring an internet connection and to share the same XR space. This allows the system to be easily introduced in a wide range of environments, including patients’ homes, nursing homes, and day care centers. Based on this shared XR-space system, I also developed multiple gamified rehabilitation contents for upper-limb functional training.
During the MITOU Program period, I conducted XR rehabilitation workshops using these developed applications at multiple nursing care facilities, where a total of 24 users experienced the system. As a result, the project received positive evaluations across all surveyed items, including enjoyment of XR rehabilitation, group experience, motivation for rehabilitation, comparison with conventional rehabilitation, and usability. Participants also shared encouraging comments such as, “I want to try this again!”
Creators: HIRATSUKA Shintaro
From Blank Canvas to Final Render: Modeling Creator Production Pipelines in Generative Models
We present LimeGen, a AI generative system that reproduces the production workflow human creators actually follow. The workflow proceeds from layout specification to foreground generation, background generation, and final compositing. Existing text-to-image models synthesize an entire image in a single pass, yielding a flat output whose parts cannot be revised without full regeneration. LimeGen instead generates and edits the image layer by layer. Each stage of the pipeline yields an explicit, editable intermediate. A creator can revise the foreground without disturbing the background, adjust the layout without discarding the rendered content, and recompose the result at any point.
Creators: MORITA Ryugo, NAGAI Daichi
PM: OCHIAI Yoichi
An intuitive platform for editing and sharing content in the relational context of arts and culture
The “ArsTraverse” project developed an intuitive editing platform for exhibition curators and researchers with the aim of reducing the burden of contextual research, curation, and sharing in the arts and culture sector. Previously, because much of the source material consisted of unstructured text, organizing relationships between artists, works, and concepts required a significant amount of time, and creating visualizations and explanatory content demanded a high level of expertise and incurred substantial costs. To address this, the product incorporates a feature that uses large language models (LLMs) to extract subject-predicate-object relationships from textual materials and semi-structure them into a knowledge graph. It integrates a repository that cross-references multiple sources, intuitive drag-and-drop editing, synchronization between written text and related concepts, and storytelling editing and output that maps text to partial graphs, thereby providing comprehensive support from research to dissemination. Furthermore, the system is designed to accommodate annotations, version control, and proposal approval workflows, enabling the divergent accumulation of diverse interpretations. We updated the requirements through on-site interviews and workshops, demonstrating the system’s potential for use in exhibition settings. Moving forward, we aim to advance further social implementation and, through collaboration among cultural institutions and the deployment of open-source software (OSS), develop this into a knowledge resource infrastructure for the arts.
Creator: MATSUNO Hitoshi
A synthesizer for intuitive sound design through synesthetic visualization
In short, the system developed in this project is a synthesizer that enables users to “see” timbre.
Although conventional synthesizers can produce a wide variety of timbres, users typically need to audition presets one by one to understand how they sound, and the relationship between parameters and timbre is often difficult to grasp. To address these issues, this project focuses on timbre visualization as a way to improve the intuitiveness of the user interface for sound design.
The proposed system visually represents timbre using shape and material. These visual features are estimated by a deep learning model based on each user's synesthetic associations (or cross-modal correspondences).
In the preset selection interface, the visualization result for each preset is displayed as an icon, enabling users to compare timbres at a glance before auditioning them. In the parameter editing interface, visualization results are displayed at the corresponding positions on sliders, helping users intuitively understand the relationship between parameters and timbre.
The image estimation model, software, and hardware were designed, developed, and integrated into a user-facing standalone synthesizer system.
Creator: MAIE Ayato
Development of Curve-Programmed Interlocking Structures Integrable with Sheet-Like 3D Printed Structures
The joining of 3D-printed parts is commonly achieved through methods such as thermal welding, adhesives, or simple interlocking geometries. However, these approaches present challenges in terms of assembly time, labor, and structural strength, particularly when applied to the thin sheet-like or textile-like structures frequently used in 3D printing.
In this project, we developed a method for generating interlocking structures inspired by the zippers commonly found in garments, adaptable to arbitrary target geometries. Because the interlocking mechanism is fabricated integrally with the printed parts, it enables strong assembly without requiring additional components or adhesives. Furthermore, by appropriately deforming the base structure, the system can be applied to complex curved surfaces and also enables novel textile expressions in which the material transforms through the act of assembly itself.
To propose applications in the field of fashion design, we also actively explored tools for segmenting and unfolding 3D clothing models into smaller printable parts, as well as textile designs that emphasize the waviness and deformation created along the joints. The project was presented in the form of a fashion show featuring five garments at the final presentation event.
Creator: MASUDA Ryo, NAGATA Risa, KANEDA Masaya
PM: SOGAWA Keisuke
Development of a Live-Migratable WebAssembly Runtime and a Library OS Using WASI as Its Interface
In this project, I developed "Kafu," a framework that executes WebAssembly (Wasm) modules in a distributed environment that transparently integrates the cloud and the edge, and performs inter-node migration at function-call boundaries. In recent years, the importance of cloud–edge collaboration has grown, driven by the need to analyze data from robots, drones, and other devices immediately on-device, as well as the need to perform computationally intensive processing in the cloud. However, developing conventional cloud–edge applications requires complex implementations involving network communication, RPC, and state management, resulting in high development costs.
Kafu introduces the concept of aspect-oriented programming and enables developers to dynamically specify execution nodes simply by adding function attributes (annotations) to C/C++ source code. This allows developers to describe distributed services as if they were writing a single-process application, without needing to explicitly handle network communication or state persistence. I adopted WebAssembly as the foundation for distributed execution because of its high portability and lightweight nature.
In addition, I developed a live migration technique for WebAssembly runtimes and implemented (Snapify), a novel binary transformation method that allows programs to be suspended and resumed without depending on a specific runtime implementation. By accelerating communication through differential memory transfer and compression, I demonstrated that live migration can be completed in approximately 100 ms for a real-time image processing program. I distributed the resulting system as an SDK together with an installer, and released the source code on GitHub under an open-source license.
Creators: TAMURA Raiki
Document Editing System for Advanced Typesetting with Intuitive Interfaces
Today's document creation systems are largely divided into two categories: "Microsoft Word," which excels in intuitive usability, and "LaTeX," which specializes in beautiful typesetting. These two approaches have historically presented a trade-off. This project resolves the challenges of both and has developed "Monk," a document creation system that achieves both the intuitive usability of Word and the advanced typesetting of LaTeX.
The prominent feature of Monk is the complete real-time bidirectional integration on the screen between the "CUI," where users write the Monk language, and the "GUI," which displays the final rendered output. Users can describe complex documents precisely from the CUI in Monk’s text-based language, while also editing the rendered result directly through the GUI. When a user modifies the visual document, the corresponding source code is updated automatically.
To make this possible, we designed a new statically typed language that combines the readability of markup with the expressiveness and flexibility of programming languages. The language helps users write clear, maintainable documents while reducing errors. Furthermore, it complies with standard typesetting requirements, allowing for the beautiful arrangement of text, including mathematical formulas.
The greatest advantage of this system is that users can choose to operate either the CUI or the GUI depending on what suits their situation best. Beginners can significantly lower the barrier to entry by starting with the intuitive GUI. At the same time, advanced users can speed up their writing by choosing the more suitable interface for their current needs. The completely new value of this system lies in its ability to achieve beautiful typesetting while providing an easy and comfortable user experience.
Creators: MARUYAMA Takara, ISHIDA Tempo, NAKAMURA Ken
A Go Framework for Code-Based Software Design and Formal Verification
In this project, we developed goat, a framework that enables software design to be written and verified as code, and YAGI, an agent application that generates code-based designs through interaction with AI. With goat, designs can be expressed in Go code and mechanically verified using model checking. This allows the correctness of a design to be validated before implementation, reducing rework and failures caused by design flaws. In addition, goat can generate interface definitions and E2E test code from the design, preventing the separation between design and implementation and enabling consistent development.
Meanwhile, AI-assisted development today often relies on ambiguous natural language instructions, which can lead to discrepancies between intended and generated code. By using YAGI, developers can generate designs in a code-based, verifiable form through interaction with AI, and provide them directly to coding agents. This makes it possible to communicate intent in a precise and verifiable manner, reducing ambiguity in interpretation and enabling more accurate and reproducible specification-driven development.
Creators: TODA Honoka
PM: TAKESAKO Yoshinori
Development of an Operating Platform Integrating OS and Web Browsers Powered by Local LLMs
Today’s web browsers handle a wide range of tasks, yet they remain isolated from the local OS, resulting in inefficiencies. Furthermore, high-performance cloud-baed LLMs face challenges regarding confidentiality and cost. This project aims to develop “Floorp OS,” a next-generation LLM platform that eliminates the boundary between the web and the local OS, with the highly customizable Floorp web browser at its core. Specifically, we have developed “Sapphillon,” a foundation that generates explicit workflows from natural language instructions and automatically executes them as a series of processes. This foundation operates in a sandbox environment and features granular permission management and an extensible plugin system. Furthermore, by leveraging a local LLM, secure processing is possible without sending confidential data externally. This enables users to safely and efficiently automate tasks that span both website operations and local file processing. This development is expected to be utilized by companies and research institutions that have found it difficult to adopt cloud-based AI due to information management constraints, thereby expanding the possibilities for new user experiences and practical automation in everyday PC operations.
Creator: TAKAHASHI Yuta, ASANO Ryosuke, AYUZAWA Hayato
Development of a 3D Printing Preprocessor for Robot Contests
This project addressed the critical challenge of balancing lightweight design with high structural strength in manufacturing fields like robotics competitions by developing "Strecs3D," a software tool that locally optimizes 3D-printed internal (infill) structures based on stress distribution. Conventional 3D printing typically applies a uniform infill density throughout a part, resulting in structural inefficiency where material is wasted in low-load areas while strength remains insufficient at points of stress concentration. To resolve this, I proposed a unique workflow that introduces a "slicing preprocessor" layer between design (CAD) and manufacturing (slicer). The system performs Finite Element Analysis (FEA) on the 3D model to calculate stress distribution and automatically partitions the model into optimal subdomains with varying densities calculated via the Gibson-Ashby model. This optimization data is integrated as manufacturing metadata into a 3MF file, enabling seamless collaboration with major existing slicing software. Experimental evaluations demonstrated that compared to standard uniform infill models of the same mass, Strecs3D achieved approximately a 1.97-fold increase in maximum load capacity and a 1.65-fold increase in flexural rigidity. Ultimately, this project establishes a highly efficient digital manufacturing pipeline grounded in quantitative mechanical evidence rather than empirical intuition.
Creators: TANIGUCHI Tomohiro
A Programming Assignment Grading System Based on Collaboration among AI, Learners, and Educators
In this project, we developed a new grading system that combines LLM-based automatic code evaluation for programming assignments with crowdsourcing-style peer assessment among learners. Conventional automatic grading systems mainly determine correctness based on standard output, and have not been sufficient for evaluating visual and dynamic programs. Manual grading also becomes increasingly burdensome for instructors and TAs as the number of students grows, making it difficult to provide prompt feedback.
In this system, when learners submit an assignment, their code is first automatically evaluated by an LLM. In addition, the system incorporates a process in which learners evaluate the execution results of other students’ submissions while interacting with them. This distributes the grading workload and reduces the burden on instructors. Through this approach, learners can notice ambiguities or shortcomings in the assignment specifications while evaluating others’ submissions, giving them opportunities to revise their own work. Furthermore, by distributing the grading process in advance, the system shortens the waiting time for feedback and enables a smoother learning cycle.
Through this project, we improved the efficiency of evaluating dynamic programs, reduced the workload associated with grading in educational settings, and created an environment in which learners can engage in more self-directed learning.
Creator: SEKIGUCHI Yuto
PM: TANAKA Kunihiro
Portable Shape Display Using an Ultra-Compact Pin Array
This project focuses on pin-array devices that present shapes through the vertical motion of pins, and aims to realize a portable shape display that achieves both high density and portability.
Shape display devices have been studied as technologies for presenting visual information through the sense of touch, particularly in fields such as Braille displays and tactile interfaces. However, existing devices face limitations in the number and resolution of pins, making it difficult to achieve high-density shape presentation with sufficient expressive capability.
A major factor behind this limitation is the structural constraint that arises as the number of pins increases. Increasing the number of pins requires a corresponding number of actuators to drive them. In conventional approaches, each motor requires its own wiring and control circuitry, causing the circuit scale and wiring complexity to increase in proportion to the number of pins. As a result, the device becomes larger, while power consumption and cost also increase, making it difficult to maintain portability.
To address these challenges, this project designs both the high-density pin arrangement and the driving method that supports it. A key feature of the project is that it considers mechanical design and circuit design in an integrated manner.
Ultimately, the project aims to combine a structure in which many pins are densely arranged with an efficient circuit architecture for controlling them, and to build a working prototype of a shape display device.
Creator: SUGIMOTO Ryuhei
Development of a High-Performance, Fault-Tolerant MySQL-Based System
In this project, we developed Kamo, a fault-tolerant open-source database management system that delivers higher performance than MySQL. By adopting the high-performance LineairDB in MySQL’s pluggable storage engine layer, Kamo significantly improves performance while maintaining compatibility with MySQL. Furthermore, by implementing an automatic failover mechanism that detects node failures in a replication environment and promotes a replica to the new primary, Kamo achieves fault tolerance that enables services to continue operating even when failures occur. In this project, the developed Kamo achieved an 18.7× performance improvement under the TPC-C workload with 144 threads, while also realizing effectively zero seconds of downtime during failures through a design that integrates Orchestrator and ProxySQL.
Creator: MIYAZAKI Yusuke, NAKAMORI Tatsuhiro, RI Hirofumi
Exoskeleton with a new mechanism and its ergonomic design
In this project, we developed Mechanical Suits, a walking-assist suit built around a mechanism that uses non-circular gears. A defining feature of the design is that it contains no powered components such as batteries or motors. The non-circular gear mechanism modulates the output of internal springs, delivering powerful assistance synchronized with the user's natural gait.
In addition, we developed automated design software. By inputting parameters such as height, weight, and walking speed, the system automatically generates a custom suit design tailored to the individual wearer. Because the suit's primary components are fabricated with 3D printers, the entire workflow from personalized design to custom manufacturing is fully automated. This personalization-driven design approach enables the suit to achieve high performance despite being unpowered.
Furthermore, the production cost is remarkably low—on the order of only a few thousand yen—providing a substantial economic advantage over conventional assist suits, which typically fall in a much higher price range.
Creators: MURAKI Yuta
Change log
-
Sep 30, 2026
Released this page