Nozomi Akashi

Cellulose nanofibers (CNFs) are attractive materials for green electronics, but their use as the computational core of intelligent devices remains largely unexplored. Here, a hydrogel made of CNFs and water is examined as a physical computational resource, i.e., a physical reservoir, for electronic time‐series processing, based on a square‐gel measurement platform with three identical input and 13 readout electrodes. The dynamics depend on pulse width, where a slower relaxation branch becoming increasingly relevant at longer pulses. Increasing pulse width shifts the reservoir from a linear‐memory‐rich regime toward a more contractive and more nonlinear regime, where more contractive means that the same input drive suppresses differences between neighboring states more strongly. Useful functionality is retained for more than 3.5 h after being exposed to an ambient drying environment: sub‐millisecond conditions favor a benchmark for nonlinear memory capability (NARMA2), whereas node emulation, the restoration of outputs from a missing node by available ones, is advantageous for longer pulse widths. These results establish bare CNF hydrogel as a degradable physical reservoir with drying robustness.

Reservoir computing (RC) is a machine learning framework that uses recurrent neural networks and is characterized by directly capitalizing on intrinsic dynamics instead of adjusting internal parameters. In particular, in the form of physical reservoir computing (PRC), recent studies have advanced by treating various physical systems as reservoirs and applying them to time-series data processing and quantifying information-processing properties. In this way, RC and PRC potentially have interdisciplinary impact, and as more researchers from diverse academic disciplines learn and utilize RC and PRC, there is potential for more creative research to emerge. In this paper, we introduce a Jupyter Notebook-based educational material called RC bootcamp for learning RC, being made publicly available under an open-source license (https://rc-bootcamp.github.io/). The RC bootcamp was originally developed and continuously updated within our research group to efficiently train our collaborators and new students, ultimately enabling them to conduct experiments by themselves. Considering the diverse backgrounds of learners, it starts with the basics of computer science and numerical computation using Python/NumPy, as well as fundamental implementations in RC, such as echo state networks and linear regression. Furthermore, it covers important analytical indicators based on dynamical systems theory, such as Lyapunov exponents, echo state property index, and information-processing capacity, as well as cutting-edge approaches utilizing chaos, including first-order, reduced and controlled error (FORCE) learning and innate training, and attractor design via bifurcation embedding. We expect that the RC bootcamp will become a useful educational material for learning RC and PRC and further invigorate research activities in the RC and PRC fields.

Understanding and predicting how mechanical systems respond to environmental variability is essential for advancing next-generation robotic systems with physical intelligence. In this study, we investigated the use of echo state networks (ESNs), a representative class of reservoir computing (RC) models, to predict the bifurcation structures of real-world mechanical systems from limited observations. We examined two representative cases: a simulated passive dynamic walking (PDW) robot with hybrid continuous–discrete dynamics and a real-world soft pneumatic artificial muscle (PAM) actuator whose electrical resistance undergoes complex changes under varying loads. To address the challenges posed by the PDW’s hybrid dynamics, we proposed a hybrid ESN (HESN) model that integrates a knowledge-based touchdown detection mechanism with an ESN module. The HESN successfully reproduced the route-to-chaos bifurcation structure of the PDW, captured its multi-attractor dynamics, and demonstrated robustness against imperfect domain knowledge. For the PAM, where no reliable physical model is available, a purely data-driven ESN accurately predicted resistance bifurcations across changing environmental conditions. These results highlight the potential of RC models as flexible digital twins for mechanical systems, enabling parameter-aware modeling of bifurcations with limited training data and supporting the design of robust, adaptive robots capable of operating in complex environments.

Machine learning methods have shown promise in learning chaotic dynamical systems, enabling model-free short-term prediction and attractor reconstruction. However, when applied to large-scale, spatiotemporally chaotic systems, purely data-driven machine learning methods often suffer from inefficiencies, as they require a large learning model size and a massive amount of training data to achieve acceptable performance. To address this challenge, we incorporate the spatial coupling structure of the target system as an inductive bias in the network design. Specifically, we introduce physics-guided clustered echo state networks, leveraging the efficiency of the echo state networks (ESNs) as a base model. Experimental results on benchmark chaotic systems demonstrate that our physics-informed method outperforms existing echo state network models in learning the target chaotic systems. Additionally, we numerically demonstrate that leveraging coupling knowledge into ESN models can enhance their robustness to variations of training and target system conditions. We further show that our proposed model remains effective even when the coupling knowledge is imperfect or extracted directly from time series data. We believe that this approach has the potential to enhance other machine learning methods.

Chaos presents complex dynamics arising from nonlinearity and a sensitivity to initial states. These characteristics suggest a depth of expressivity that underscores their potential for advanced computational applications. However, strategies to effectively exploit chaotic dynamics for information processing have largely remained elusive. In this study, we reveal that the essence of chaos can be found in various state-of-the-art deep neural networks. Drawing inspiration from this revelation, we propose a novel method that directly leverages chaotic dynamics for deep learning architectures. Our approach is systematically evaluated across distinct chaotic systems. In all instances, our framework presents superior results to conventional deep neural networks in terms of accuracy, convergence speed, and efficiency. Furthermore, we find an active role of transient chaos formation in our framework. Collectively, this study opens up a new path for the integration of chaos, which has long been overlooked in information processing, and provides insights into the potential fusion of chaotic dynamics within the domains of machine learning and neuromorphic computation.

The optimization of physical parameters serves various purposes in device development, including in system identification and efficiency. Spin-torque oscillators have been experimentally and theoretically applied to neuromorphic computing, but their physical parameters are usually optimized via grid search procedures. In this paper, we propose a scheme to optimize the dynamics parameters of macrospin-type spin-torque oscillators using the gradient descent method with automatic differentiation. First, we numerically create dynamic data for teaching and tune the parameters to reproduce the dynamics. This approach can be applied to determine the correspondence between spin-torque oscillator simulations and experiments. Next, we solve an image recognition task with high accuracy by connecting a coupled system of spin-torque oscillators to the input and output layers and training all of them through gradient descent. Combining this approach with experimentation makes it possible to design an experimental setup and physical system to solve a task with high precision using a spin-torque oscillator.

Harnessing complex body dynamics has long been a challenge in robotics, particularly when dealing with soft dynamics, which exhibit high complexity in interacting with the environment. Recent studies indicate that these dynamics can be used as a computational resource, exemplified by the McKibben pneumatic artificial muscle, a common soft actuator. This study demonstrates that bifurcations, including periodic and chaotic dynamics, can be embedded into the pneumatic artificial muscle, with the entire bifurcation structure using the framework of physical reservoir computing. These results suggest that dynamics not present in training data can be embedded through bifurcation embedment, implying the capability to incorporate various qualitatively different patterns into pneumatic artificial muscle without the need to design and learn all required patterns explicitly. Thus, this study introduces a novel approach to simplify robotic devices and control training by reducing reliance on external pattern generators and the amount and types of training data needed for control.

A passive dynamic walker is a mechanical system that walks down a slope without any control, and gives useful insights into the dynamic mechanism of stable walking. This system shows specific attractor characteristics depending on the slope angle due to nonlinear dynamics, such as period-doubling to chaos and its disappearance by a boundary crisis. However, it remains unclear what happens to the basin of attraction. In our previous studies, we showed that a fractal basin of attraction is generated using a simple model over a critical slope angle by iteratively applying the inverse image of the Poincaré map, which has stretching and bending effects. In the present study, we show that the size and fractality of the basin of attraction sharply change many times by changing the slope angle. Furthermore, we improved our previous analysis to clarify the mechanisms for these changes and the disappearance of the basin of attraction based on the stretching and bending deformation in the basin formation process. These findings will improve our understanding of the governing dynamics to generate the basin of attraction in walking.