Results of the FY2025 MITOU Target Program: Software Development Utilizing Reservoir Computing Technology
Release Date:Sep 30, 2026
In-page Links
Results of the FY2025 MITOU Target Program: Software Development Utilizing Reservoir Computing Technology
PM: KATORI Yuichi
Development of Reservoir Computing Software for Real-Time Robot Learning
This project developed a software framework that enables high-speed online learning based on Reservoir Computing (RC) for the real-time control of multi-degree-of-freedom (DOF) robots. In conventional Python-based RC implementations, the computational bottleneck in Recursive Least Squares (RLS) updates becomes severe as the reservoir size scales up. This bottleneck makes it difficult to embed learning mechanisms within the real-time control loops. To overcome this, a novel RC library was developed that optimizes the core computational algorithms in C++ while providing a user-friendly Python interface, achieving both high execution speed and practical convenience. The library was integrated into the control system of a 7-DOF robot arm, where a teaching-playback demonstration and adaptive gravity compensation for dynamic load variations were successfully implemented. Hardware experiments confirmed that the system adapted to load variations in real time under a 50 Hz control cycle. The software has been released as an open-source library. By offering an accessible Python interface with a high-speed C++ backend, this project lowers the barrier to adopting RC techniques within the machine learning community and establishes a foundation for real-world applications in robotics.
Creator: ATSUTA Hiroshi
Next-Generation Tactile Interfaces Enabled by Adaptive Reservoir Computing
To realize tactile AI that enables robots to change their behavior according to the situation based on tactile information, we developed an autonomous exploratory touch system centered on reservoir computing (RC). A robot arm equipped with tactile and force sensors acquires time-series data from material surfaces, and RC is used to perform material classification and confidence estimation. When the confidence does not reach a threshold, the classification result is not adopted; instead, Bayesian optimization updates the tracing direction, speed, and distance, and the system remeasures the surface to stabilize the decision. Training and inference were conducted on four types of materials, confirming that correct classification was achieved once the confidence became sufficiently high. In addition, we implemented the RC reservoir layer on an FPGA and demonstrated behavior equivalent to the software implementation, as well as real-time performance of 1.2 ms per sample, compatible with 600 Hz measurement.
Creators: TAKESADA Kazuki, MIKI Katsuto
NeuMoReservoir: Spatiotemporal Dynamics Emerging from Neuronal Morphology
I developed “NeuMoReservoir,” a software tool for evaluating the computational capabilities emerged from the complex morphology of neurons from the perspective of reservoir computing. Based on Python and the NEURON simulator, it treats a single-neuron model as a physical reservoir, enabling quantitative performance evaluation while varying cell morphology, synaptic placement on dendrites, time constants, and other parameters. It supports tasks such as time-series prediction, random spike train classification, and speech recognition, and also includes analytical tools such as effective rank and principal component analysis. In spoken digit recognition, it achieved a maximum accuracy of 70.4%, demonstrating that the spatial morphology itself strongly influences computational performance. Furthermore, these results suggest that efficient computation can be achieved without the need for neuronal networks.
Creators: FUKAMI Satoshi
PM: KAWAI Yuji
Development of an Interactive Reservoir Computing Demo for Learning and Outreach
I developed “Reservoir Play,” a website for hands-on learning of reservoir computing, and created an environment in which everything from training to real-time inference can be completed within a web browser. I implemented demos for speech classification and gesture classification that run entirely in JavaScript, allowing users to try them with no installation required. In addition, I presented an edge AI implementation example in which a model trained on a PC is deployed to a micro:bit, demonstrating that inference is possible even under limited ROM and RAM resources. This work turns a previously theory-centered technology into practical educational material that can be understood by actually running it, and establishes a foundation for its use in education, hands-on workshops, and embedded and manufacturing applications.
Creators: UEDA Hiroshi
Development of Quantum Phase Hybrid Reservoir Computing
I developed “quantum phase hybrid reservoir computing,” a framework that classically couples multiple quantum systems belonging to different quantum phases to design computational performance. In conventional quantum reservoir computing, performance improvement typically requires precise tuning of quantum parameters, which imposes a substantial implementation burden. In the proposed method, each quantum system receives an input and undergoes time evolution, measurement, and linear readout computation; the measured outputs are then mixed with the original input and fed back into the systems, thereby forming a dynamical phase in which multiple quantum phases interact. In a delayed trigonometric product emulation task, a hybrid configuration combining a quantum chaotic phase, which excels in nonlinearity, and an integrable phase, which excels in memory, achieved the best performance under high-frequency and long-delay conditions, confirming that the advantages of both phases can be combined. By treating the combination of quantum systems itself as a new design degree of freedom, this framework enables practical performance tuning while maintaining stability and reproducibility.
Creators: KOBAYASHI Kaito
Development of a Brain Organoid Design Support Simulator Using Reservoir Computing
I developed a Python-based simulator that uses reservoir computing modeled on the spontaneous activity of brain organoids to evaluate in advance how culture conditions and network structure affect computational performance. The simulator takes MEA spike time-series data as input and reproduces the connection structure of multiple organoids using Wilson–Cowan models arranged on a 4×4 grid; it can also estimate coupling weights and time constants through Bayesian optimization. Evaluation using NARMA10 and SinSum showed that performance degrades as spontaneous activity noise increases: for NARMA10, the NRMSE was 0.136 without noise and 0.184 with biologically comparable noise. By quantitatively comparing computational performance in systems that include biologically derived noise, this work showed that controlling spontaneous activity is an important requirement in the design of biological reservoirs.
Creators: FUJIMOTO Asato
PM: TANAKA Gohei
Toward the Realization of AI Robots with Innate Individuality Using LSI Manufacturing Variations in Reservoir Computing
This work presents an FPGA-based AI model with hardware-derived “innate individuality,” created by incorporating subtle chip-to-chip variations that arise during LSI fabrication into reservoir computing as fixed, non-updated weights. To realize this concept, we developed an ITCAM with a variation-extraction circuit and an IRC that uses the extracted values as its initial weights. We evaluated the system on dynamic XOR tasks using four Spartan-7 boards and four Cyclone IV boards. While the maximum accuracy generally reached around 90–95%, clear chip- and board-dependent differences emerged in the minimum accuracy and error patterns, causing each board to exhibit distinct behavior even under identical training conditions. These results demonstrate a technological foundation for AI robots that can exhibit consistent individuality determined by the characteristics of the chip installed in each device, even when the products themselves are otherwise identical.
Creators: OGAWA Masahiro
Development of a Non-Contact Vital Data Measurement and Analysis App for Dogs and Cats Using Reservoir Computing
I developed an iOS app that measures and analyzes the resting respiration of dogs and cats in a non-contact manner using approximately 15-second smartphone videos. The video is converted into a respiratory time series through image alignment, optical flow, noise removal, and Z-score normalization, and the respiratory rate per minute is then calculated by frequency analysis. To assess breathing patterns, I adopted an ESN-based time-series predictor. For each individual animal, a model is trained on resting-state videos, and the mean absolute prediction error is used as an anomaly score to detect changes such as temporary increases, decreases, or pauses in breathing rate. Pet information, trained weights, and measurement results are managed separately within the device, allowing the app to be used even in offline environments.
Creators: SHIBUTA Kenta
Development of Reservoir Design Support Software Using Multi-Objective Optimization
To reduce the reliance on individual experience and intuition in echo state network (ESN) hyperparameter design, we developed MORSe, a multi-objective optimization-based hyperparameter optimization (HPO) software tool. MORSe simultaneously optimized multiple metrics depending on the task, such as accuracy, error, and the number of nodes, and visualized the search results and trade-offs among these metrics as a Pareto front in a web browser. In addition, MORSe displays hyperparameter importance and reservoir network metrics to support the analysis and selection of optimization results. For the search, we employed MOEA/D and extended it to handle mixed continuous, integer, and categorical variables as well as interactions among variables. Custom tasks and custom ESN implementations could also be added by inheriting from abstract base classes. Through these capabilities, MORSe aims to reduce the burden of developing time-series processing systems and to promote fit-for-purpose AI design under resource constraints.
Creators: TOMITA Noriyuki, SUNAYAMA Yosuke
Change log
-
Sep 30, 2026
Released this page