Results of the First Term of the FY2025 MITOU Advanced Program
Release Date:Sep 30, 2026
In-page Links
Results of the FY2025 MITOU Advanced Program
PM: ISHIGURO Hiroshi
Developing a Virtual Care Companion for People Living with Dementia and a Support Platform
We developed “Tomori,” a virtual helper designed to mediate between people with dementia and their families, reducing the burden of home caregiving and the conflicts that can arise. In dementia care, support for everyday tasks such as taking medication, eating meals, drinking water, preparing to go out, and checking locks is easily interrupted, leading to caregiver exhaustion and anxiety for the person receiving care. Tomori is a stationary conversational robot powered by an LLM that provides reminders, relays messages, responds to repeated questions, and offers emotional support through casual conversation and empathy. For family members, it includes a LINE-linked user interface that allows them to send requests and check conversation logs remotely. Its key features are ease of use—simply set it up and turn it on—and an appearance and dialogue design that make it easy to accept in the home. Demonstration trials also confirmed cases in which prompts from the robot were more effective than prompts from family members in encouraging action. We are advancing commercialization through ReMENTIA and aiming to provide a beta version and promote adoption in collaboration with local governments.
Creators: MIYASHITA Takuma, GOTO Taisei, HONDA Junya, YAMAMOTO Kento
PM: URUSHIBARA Shigeru
Digital Applications for Skill Enhancement and Technique Transmission in the Traditional Craft of Kumiko
Traditional woodworking craft known as “kumiko” relies heavily on artisans’ experience and intuition in both design and production, making it difficult to learn and pass on. To address this challenge, we developed “kumikoAI,” a co-creative design system that can generate kumiko designs from images or text and supports editing, shading adjustment, layer visualization, and even verification with 3D models. By enabling artisans to refine AI-generated proposals and deepen designs through repeated back-and-forth between digital design and physical production, the system helps formalize tacit craft knowledge while opening up new forms of expression. In addition to conventional flat patterns, we also proposed design and production methods for three-dimensional kumiko with relief and depth, leading to a patent application. The system has already been applied in municipal projects and Expo-related works, demonstrating its potential to support the transmission of traditional craft skills, encourage participation by younger and cross-disciplinary creators, and promote regional culture both domestically and internationally.
Creators: ISHIMOTO Daiho, MAE Narumi, YAMAZAKI Shodai
The Development of a Next-Generation, Web-Based CAD Platform via Procedural Modeling
In additive manufacturing, complex designs such as lattice structures, porous structures, and internal channels are increasingly important because the geometry itself determines functionality. However, with the explicit geometric representations commonly used in conventional CAD, operations such as Boolean computation and thickening are prone to failure, making stable design difficult. To address this issue, I developed “Nodi,” a CAD platform that integrates explicit geometry and distance-function-based implicit geometry within a single CAD kernel and enables them to be handled consistently through node-based procedural modeling. Nodi includes a lightweight browser-based version and a desktop version designed for computationally intensive tasks and stricter security requirements, both of which share the same operational framework. Another key feature is its high portability through the use of Rust and WebAssembly, making implicit modeling—previously requiring specialized expertise—more visual, accessible, and reusable. During the project period, Nodi secured two domestic orders and one international order. It is expected to promote design methodologies tailored to additive manufacturing, strengthen the linkage between design and production, and support the global expansion of a design platform originating in Japan.
Creators: NAKAMURA Masatatsu
PM: KAJITA Mami
Development of a Cloud-integrated Map SDK Unification Platform
In mobile map application development, each SDK, such as Google Maps and MapKit, has different APIs, performance characteristics, and pricing models, making migration and optimization costly. To address this issue, we developed “MapConductor,” a middleware SDK for Android and iOS that enables multiple map SDKs to be handled through a unified API, allowing developers to describe markers, polylines, polygons, clustering, heatmaps, and other map elements with the same syntax and behavior. A key feature is its support for declarative UI, which allows developers to focus on map representation without being concerned with differences between operating systems or map SDKs. We also prototyped an adapter for ArcGIS Online integration and a location information provider that combines GPS and accelerometer data to reduce power consumption. Through proprietary logic that minimizes SDK dependency, we achieved support for many map SDKs in a short period of time, and as an open-source software project, we aim to build a community and democratize map application development.
Creators: OHMORI Takaaki, KATSUMATA Masashi
AI Agent-Integrated Remote Medical Support System
Multimorbidity often requires clinical decision-making across specialty boundaries, while the uneven distribution of specialists can limit timely access to the expertise needed at the point of care. To address this challenge, we developed Doctors Summit, a chat-based physician-to-physician teleconsultation system incorporating an AI agent. The system manages case-based discussions, structures free-text clinical information into a format useful for clinical decision-making, identifies missing information, and supports the initial organization of consultation content to improve consultation quality. In a pilot involving approximately 40 physicians across 20 institutions, the system supported 23 consultations over 1.5 months; 93% were resolved through remote specialist advice alone, and AI-assisted structuring of consultation records was estimated to reduce physicians’ documentation time by 65%. With compliance with relevant medical regulations as a prerequisite, we are advancing real-world implementation through collaborations with academic hospitals and industry partners, expansion of the physician user base, and clinical research.
Creators: SAWANO Shinnosuke, TAKEUCHI Hirotoshi, CAMERON Thomas Kenta
PM: SHUDO Kazuyuki
Development of a Compact Sensor for Taste and Quality Evaluation Using Electrical Impedance Properties
In restaurants and food service settings, taste can easily vary due to differences in ingredients, temperature, and worker skill, and final judgment often depends on the intuition of experienced staff. To address this issue, we developed a compact taste and quality evaluation sensor that applies a minute alternating current to liquid foods and beverages and captures taste as a multidimensional “state” arising from the interaction of multiple components, based on multi-frequency electrical impedance properties. Unlike single-parameter measurements such as pH, sugar content, or salinity, the sensor can evaluate subtle differences in taste in a multidimensional manner, and the entire process from measurement to result display can be completed on the device alone. In addition, by optimizing electrode placement, signal paths, and drive/acquisition methods, we ensured reproducibility without relying on software-based correction, enabling stable operation even under the variable conditions of real cooking environments. The sensor is expected to be applied to quality control in restaurant chains and food manufacturing, as well as to reducing the need for re-cooking, minimizing waste, and lowering training burdens.
Creators: KOBAYASHI Akihito, KIMURA Yuki
Development of a System to Prevent Phone-Based Fraud
Phone-based fraud targeting older adults often begins with incoming calls to landlines or smartphones. To prevent such scams at the point of contact without requiring users to change how they use their phones, we developed “Sagi-dome Taro,” a real-time phone scam prevention system.
The landline version analyzes call audio in real time and, when a potential scam is detected, alerts the user through voice, visual, and vibration warnings, prompting them to end the call. By adopting an RJ9 connection that plugs directly into the handset cord and a dual-threshold detection mechanism, the system enables continuous automatic monitoring while reducing noise-induced false alarms. We also reduced the hardware cost from approximately JPY 10,000 to less than JPY 4,000.
The smartphone version addresses OS-level restrictions on accessing call audio by combining external hardware with a proprietary communication method, enabling analysis and intervention from the beginning of an incoming call.
We are also conducting demonstration trials and hands-on events in collaboration with police departments and local governments. Through the development of an LTE-M-enabled version and preparation for mass production, we aim to establish a scalable phone scam prevention infrastructure that can be widely deployed without requiring users to configure or change how they use their phones.
Creators: NISHITANI Sotetsu, TAIRA Humiya, OGATA Koutaro, OJIMA Mutsuki
Development of a Next-Generation Workflow Engine for Agent Collaboration
In real-world AI agent deployment, there is still a lack of infrastructure that combines flexible, dynamic control with the robustness required for production use. To address this challenge, we developed **Graflow**, a next-generation workflow engine that combines agents capable of dynamic decision-making with the reliability of conventional workflows. Graflow provides a Python-based DSL and supports runtime generation of nodes and edges, distributed execution, type-safe state sharing, checkpointing and resumption, human-in-the-loop workflows, fine-grained error handling, observability, and integrations with a wide range of LLMs. By combining the strengths of DAGs and state machines, Graflow enables the flexible and reliable automation of complex business processes, such as responding to RFPs, reviewing contracts, and preparing materials for sales meetings. Graflow is released as open-source software under the Apache 2.0 license, with the goal of driving adoption through community development and support for enterprise deployment.
Creators: YUI Makoto, WATANABE Hiroyuki
PM: HARADA Tatsuya
Development of a Coffee-bean Analysis System to Enhance Baristas' Sampling Operations
Coffee bean sampling is a highly skilled task in which post-roast flavor is assessed through sensory evaluation, but it depends heavily on knowledge and individual perception, making it difficult to achieve consistent evaluations. To address this issue, we developed “KoScope,” a professional measurement device that combines inexpensive semiconductor-based aroma sensors with machine learning to reproduce the judgments of skilled baristas and instantly evaluate quality scores and flavor characteristics. In addition, we developed “yomnom,” a consumer-oriented coffee pot that presents flavor in written descriptions based on the aroma of brewed coffee as well as contextual information such as weather and time of day, enabling drinkers to deepen their appreciation through language and accumulate those impressions as records. The novelty of this work lies in directly converting aroma into data and applying it to both quality control and the drinking experience. While we are advancing demonstrations within the coffee industry and pre-order sales, we ultimately aim to expand this into a broader platform for treating aroma itself as information.
Creators: OKADA Takuma, HASEGAWA Taito, SUZUKI Kodai
Development of a Home-Based Pleural Effusion Assessment System to Prevent Heart Failure Exacerbation
Heart failure is characterized by frequent exacerbations and hospital readmissions, and the difficulty of objectively assessing a patient’s condition at home remains a major challenge. To address this issue, we developed a home-based pleural effusion assessment system that combines non-invasive transthoracic impedance measurement with machine learning, enabling objective assessment of intrathoracic fluid status at home or in care facilities.
The all-in-one device, “medHeart Dock,” developed by adapting existing equipment, integrates impedance measurement, machine-learning inference, and result display into a single unit. An electrode-placement guide and cloud connectivity are incorporated to facilitate simple and reliable self-measurement by patients.
A machine-learning model integrating multi-frequency thoracic impedance measurements with patient background information achieved an accuracy of 91.6% under conditions designed to simulate real-world use.
This system may facilitate self-management after hospital discharge, promote earlier medical consultation when signs of worsening heart failure are detected, and ultimately contribute to reducing hospital readmissions and healthcare costs. Furthermore, it has the potential to serve as a platform for remote patient monitoring and the real-world implementation of medical artificial intelligence.
Creator: NOSE Daisuke, MATSUDA Yuki, MATSUI Tomokazu, INOKUCHI Shoichiro, YASUMOTO Keiichi
PM: HIRANO Yutaka
Adaptive Behavior Acquisition System for a Large Hexapod Robot
Against the backdrop of Japan’s slower progress in robot commercialization compared with the United States and China, we developed “Halmonia Compass,” one of the world’s largest hexapod robots using QDD actuators, as a platform for rapidly creating custom-built large-scale robots for entertainment and performance applications that are difficult to realize by repurposing mass-produced machines. Measuring about 2 meters in length, 1.7 meters in height, and weighing about 70 kilograms, it combines low-cost, highly responsive QDD motors with the control software “V-Sido” to achieve agile and smooth motion despite its large size. It adopts a modular structure based on commonly available aluminum frames, making it easy to modify the configuration according to production needs, such as replacing legs with arms or adding a tail, while also providing high water and dust resistance. It supports AI-driven interactions, such as making eye contact and gesturing toward multiple visitors, and during its two-week exhibition at Expo 2025 Osaka, Kansai, Japan it attracted more than 130,000 visitors. In addition to applications in advertising, commercial facilities, and theme parks, Halmonia Compass is also expected to find applications in work support on rough terrain and in imitation learning research.
Creators: OSHIMA Yukako, OSHIMA Yuji
An Embroidery Design Platform Based on the Computational Design of Stitch Patterns
In embroidery production, the final result is heavily affected by factors such as thread thickness, fabric stretch, stitch order, density, and whether underlay stitching is used. At the same time, designing embroidery data requires specialized software and expertise, and it is often difficult for clients, designers, and factories to share a common image of the finished product or communicate revisions clearly. To address this issue, we developed “nuimie,” an embroidery design platform that integrates image editing, embroidery generation, previewing, and sharing in a web browser. The platform organizes images by region, reduces colors for embroidery use while preserving important elements and removing unnecessary detail, and also uses an LLM to suggest stitching parameter candidates from natural language expressions such as “rough” or “leather-like.” It further supports simulations of color order and stitch order, comment-based review, and input/output in existing file formats. By reducing the trial-and-error burden for beginners and minimizing rework in production settings, we aim to broaden the use of embroidery from individual creators to small businesses.
Creators: SHINODA Kazuhiro, HIRABAYASHI Haruma, MASEKI Tatsuya, MIYAZAKI Kakeru
PM: FUJII Akihito
Genuine Social Networking Experiences in the Post-Media SNS Era and Leveraging AI through First-Party Data Infrastructure
As social networking services increasingly function as media platforms, growing pressure on young people to curate their image and stress over posting have become major issues. To address this, we developed “Snipe,” an SNS dedicated to candid photos taken by friends, where only such photos can be posted in real time. Using facial recognition, the system automatically posts photos in which the user appears, and because users do not select the images themselves, it helps reveal natural charm and the multifaceted nature of people without excessive editing or self-staging. Its differentiation also lies in profiles composed entirely of friend-taken photos and in a user flow designed through widgets that do not support video playback. During development, we identified and refined the features most effective for continued use, turning DAU growth from an initial decline into an increase of 300 users per month. We first aim to expand the user base as a free app and eventually pursue an advertising-based revenue model. The platform is also expected to encourage more natural sharing among friends and to reconsider an SNS experience that often becomes overly centered on comparison and the need for approval.
Creators: GUNJI Daiki, KONDO Yuki, SASAKI Yudai
A Cloud Service Specialized in WebAssembly with On-premises Integration
Deploying and continuously updating AI models in a lightweight manner on resource-constrained Edge AI devices has been difficult. To address this issue, we developed “Pipit,” an edge AI execution platform specialized for WebAssembly. It is composed of three components: “waiot,” an execution environment for microcontrollers; “Mewz,” an execution environment for servers; and an orchestrator that centrally manages both. The platform enables safe updates of only Wasm applications without rewriting the entire firmware, and it supports task switching on the order of seconds without requiring a reboot. Compared with container-based approaches, it significantly reduces distribution size, and it can centrally manage environments ranging from microcontrollers to Linux servers on a Kubernetes basis. We plan to release “waiot” and the orchestrator as open-source software, while aiming to deploy “Mewz” in practical applications for telecommunications carriers and CDN providers.
Creator: UEDA Soichiro, NOZAKI Ai
PM: MIKI Hirofumi
Automated Solar Power System Design from Architectural Drawings
Designing rooftop solar systems requires creating CAD layouts tailored to the unique roof shape of each building, and including redesign work, a single project often takes more than an hour. To address this issue, we developed a web service that reconstructs the three-dimensional shape of a house from architectural drawings and automates the design of photovoltaic systems, making it applicable even to newly built or relatively new properties that are difficult to handle using satellite data alone. The service recognizes floor plans and elevation drawings through image processing, generates roof geometry using geometric computation based on an extended straight skeleton, and further integrates automatic panel placement, wiring design, shadow and reflection simulation, and economic impact estimation. Its strengths lie in design support tailored to Japanese residential structures and electrical design practices, as well as interactive editability on 3D models, contributing to greater efficiency in quotation and proposal work for solar sales businesses regardless of whether the property is new or existing. In the future, it is also expected to be applied to non-residential buildings and agrivoltaic systems.
Creators: USUI Mitsuki, TAKEUCHI Seiichiro
Development of an Automated Penetration Testing System Using AI Agents
Dynamic application security testing (DAST) tools usually rely on pattern matching, which makes it difficult to validate vulnerabilities that depend on application-specific context. Manual penetration testing can handle these cases, but it is expensive and difficult to scale. We developed a multi-agent system that runs a black-box penetration test with only a target URL as input. It handles reconnaissance and vulnerability analysis, generates and verifies proof-of-concept exploits, and produces a report. The orchestrator builds a task graph and sends each task to a specialized agent. All outbound traffic passes through an L7 egress gateway that enforces scope and rate limits and records requests and responses for later verification. With a one-hour time limit per lab, the system solved 84.07% of the PortSwigger Web Security Academy labs in our benchmark without human intervention. We also field-tested it against real-world systems through vulnerability disclosure and bug bounty programs. We are now turning it into a service for continuous penetration testing, allowing specialists to focus on work that requires human judgment.
Creators: OKAMOTO Takumasa, ABE Tatsuya
PM: MURAKAMI Akiko
AI-Driven Security Assessment for Hardening AI Chatbots
In web services that incorporate LLMs, there is a gap: conventional LLM security assessment tools cannot verify whether vulnerabilities are actually triggered on the web, while existing web security testing tools are not sufficiently equipped to handle attacks conducted through LLMs. To address this issue, we developed “VulScribe,” an automated security assessment SaaS for AI chat systems. In this system, an attack AI generates malicious prompts using techniques such as multilingual inputs, homographs, and invisible characters, while a verification AI automatically determines both the risk level of the response and whether the vulnerability is actually triggered on the web. It supports 31 types of vulnerabilities, including XSS, SQL injection, SSRF, and prompt injection, covering both the OWASP Top 10 and major LLM-specific risks. The platform fully automates the process from assessment to report generation. In performance evaluations, it achieved results comparable in some areas to a practitioner with three years of experience, and it has also demonstrated concrete results, including a total of nine CVEs and top placements in international competitions. We are advancing commercialization through an enterprise SaaS and consulting model, aiming to improve the safety of AI adoption and alleviate the shortage of security assessment professionals.
Creators: TSUJI Satoki, SUGIYAMA Yuichi
Change log
-
Sep 30, 2026
Released this page