Results of the FY2024 MITOU Target Program: Software Development Utilizing Reservoir Computing Technology

In-page Links

Results of the FY2024 MITOU Target Program: Software Development Utilizing Reservoir Computing Technology

PM: KATORI Yuichi

Development of a Reservoir-Based Sensing System for Real-Time, Context-Aware Conversations in Indoor Environments

This project developed an interactive agent system for older adults, designed to enable natural, real-time conversations that adapt to indoor environmental conditions. Leveraging reservoir computing, the system dynamically adjusts speech rate, volume, and voice frequency in accordance with the auditory and cognitive characteristics of elderly users. In addition, by integrating sensor data—such as temperature, humidity, and illuminance—the agent generates contextually appropriate dialogue to support health management and daily living activities. A cloud-based hybrid architecture allows for efficient voice-data analysis and adaptive processing. Field trials conducted in nursing facilities and senior housing provided user feedback that guided further system refinement toward real-world implementation. This technology holds promise for reducing social isolation among older adults, facilitating early detection of dementia, and fostering advancements in the smart healthcare domain.

• View the details of Development of a Reservoir-Based Sensing System for Real-Time, Context-Aware Conversations in Indoor Environments (in Japanese)

Creator: IGARASHI Toshiharu

Development of AI modules using reservoir computing that expands the application of tactile information

This project developed an AI module to utilize tactile information and implemented a system capable of flipping paper and flexible materials. Tactile information, which combines temporal and spatial data, has been underutilized in conventional AI technologies. The system employs Reservoir Computing (RC) models and reinforcement learning, using data from tactile and distance sensors to control the pressure and speed of a 3D printer. The technology successfully executed flipping actions for various materials, promoting automation in manufacturing and logistics warehouses. Additionally, it holds potential applications in analyzing the texture of folk crafts and traditional artifacts. With its cost-effective design and scalability through mass production, the system is expected to contribute to strengthening Japan's supply chain and preserving traditional manual techniques for future generations.

• View the details of Development of AI modules using reservoir computing that expands the application of tactile information (in Japanese)

Creators: TSUBOKURA Sota, TAKESADA Kazuki

PM: KAWAI Yuji

Development of a myoelectric prosthetic hand control system capable of estimating finger angle and finger joint torque using reservoir computing

This project developed a myoelectric prosthetic hand control system utilizing reservoir computing technology. Conventional myoelectric prosthetics are expensive, have low adoption rates, and face challenges in learning costs and operational flexibility. By leveraging reservoir computing, this system enables low-cost, real-time estimation of continuous finger angles and joint torque values. The system learns from myoelectric signals and finger angle data, building adaptable models that accommodate individual differences and changes in electrode positioning. It supports flexible gripping motions tailored to objects and offers effortless learning suited for everyday use. The development of cost-effective, high-flexibility prosthetics is expected to improve societal accessibility and enhance the quality of life (QOL) for amputees.

• View the details of Development of a myoelectric prosthetic hand control system capable of estimating finger angle and finger joint torque using reservoir computing (in Japanese)

Creators: MIYASIRO Kouji, BANDO Saki

PM: TANAKA Gohei

Development of a data utilization platform for edge devices using reservoir computing

The "CanteenFlow Platform" is a data utilization platform for edge devices that employs reservoir computing technology. Unlike traditional cloud-dependent AI tools, it enables real-time processing at low computational costs and supports data collection and analysis even in environments with limited network access. Key features include real-time sensor data visualization using Web Bluetooth API, fast inference processing with WebAssembly, offline data storage capabilities, and data export in JSON/CSV formats. Additionally, improvements to UI/UX and a demo of a personal fitness app have broadened its applications for researchers, technicians, businesses, and general users. This platform is expected to find use in agriculture, environmental monitoring, logistics, and more, contributing to the adoption of edge AI and addressing societal challenges.

• View the details of Development of a data utilization platform for edge devices using reservoir computing (in Japanese)

Creators: KUROTAKI Yuta, HAMADA Rena

Development of a next-generation natural language processing model using reservoir computing

This project developed a lightweight Transformer model inspired by reservoir computing to reduce computational load and power consumption while achieving high-precision natural language processing. Unlike conventional Transformers, the model alternates between fixed layers (reservoir layers) and learning layers, with some layers fixed to reduce the number of learning parameters. Additionally, parameter sharing between layers further reduces memory usage and computational costs. In English-to-Japanese translation tasks, the model maintained comparable performance (BLEU score ~27–28) to conventional models, even with increased fixed layers, achieving energy efficiency and faster learning. This technology enables deployment in edge environments, disaster situations, and medical settings, contributing to reduced operational costs and environmental impact as a sustainable AI solution.

• View the details of Development of a next-generation natural language processing model using reservoir computing (in Japanese)

Creators: NAKAMURA Jin