Embedded Files
We provide informal, accessible explanations of my research for the general public and students. For more detailed information, please refer to the related research papers.
Information Processing through Physical Dynamics
Information Processing through Physical Dynamics
“What kind of research do you do?”
The researcher swirled a glass and replied:
“This is a computer!”
This is a common way that my mentor, my collaborators, and I introduce our research to the general public. When a fluid is disturbed, a complex wave pattern forms instantaneously and eventually settles back into a calm surface. Such dynamics, or motion, follow the laws governing fluid motion, and if those laws are computed on a computer, the fluid dynamics can be simulated. What this tells us is that the computational abilities of a computer, such as memory and arithmetic operations, can be translated into fluid dynamics. Then, conversely, might it be possible to translate fluid dynamics into computation?
Approaches based on this idea include computational frameworks known as the Liquid State Machine and the Echo State Network, which today are collectively referred to as reservoir computing. By perturbing a fluid (providing an input) and observing its dynamics (obtaining an output), and then applying machine learning to the resulting input-output transformation, one can harness not only fluids but also a wide variety of other dynamical systems for computation, including robotic bodies, optics, quantum systems, spintronics, ecosystems, chemical reactions, and cultured cells. Moreover, even without the aid of machine learning, we can recognize that the physical basis of computation is everywhere in physical dynamics: in dice that generate randomness, in animal bodies that generate movement, in computers themselves, and in the brain. My research focuses on this broad theme of information processing through physical dynamics and its surrounding questions.
So far, I have worked on specific topics such as random number generation using complex dynamics including chaos, the design of brain-like computing systems, and the analysis and prediction of the complex dynamics of robots. The related research fields are extremely broad, encompassing, for example, mathematics for describing dynamics, physics for analyzing their properties, information science for handling theories of information processing, robotics and materials science for understanding devices, and even biology. In practice, this research has been advanced through collaboration with researchers from multiple disciplines, including mathematics, physics, and information science, as well as with research institutes and companies that work directly with physical devices and robotic systems.
- True Random Number Generator using a Mechanical System
- Related Fields: Dynamical Systems, Nonlinear Physics
- Keywords: Chaos, Anosov System, Random Number Generation
Random numbers are indispensable in the modern world, with applications in security, ensuring fairness, and a wide range of numerical algorithms. There are many methods for random number generation, but those produced by computers are called pseudorandom numbers and, strictly speaking, do not possess ideal properties of randomness such as unpredictability and irreproducibility. For this reason, random number generation using physical devices—known as physical random number generation—has been proposed. However, it is difficult to verify theoretically whether such physical random numbers truly have the ideal properties of randomness, and in practice their randomness is often evaluated only through certain statistical tests.
In our research, we focus on Anosov systems, which have long been studied in dynamical systems theory. Anosov systems are a class of chaotic dynamical systems that can be shown to possess mathematically strong properties such as ergodicity and topological mixing, and a corresponding physical realization has been discovered in the form of a planar linkage mechanism called the triple linkage. By using this triple linkage as a random number generator, it becomes possible to construct a generator that is physical in nature while still allowing its randomness to be proven mathematically in a rigorous way. We expect this line of research to contribute both to a deeper theoretical understanding of random number generation and to the development of more reliable physical random number generators.
Related Publications
Related Publications
N. Akashi, K. Nakajima, M. Shibayama, Y. Kuniyoshi, A mechanical true random number generator, New Journal of Physics 24: 013019, 2022.
2. Fractal Analysis of the Passive Dynamic Walker
2. Fractal Analysis of the Passive Dynamic Walker
- Related Fields: Dynamical Systems, Embodied Cognitive Science, Robotics
- Keywords: Chaos, Fractal, Bifurcation Structures, Passive Dynamic Walkers, Morphological Computation
Motion generation that interacts with the environment, such as bipedal walking, is generally thought to require precise control. However, a two-legged robot known as a passive dynamic walker can, remarkably, achieve smooth bipedal walking simply by being placed on a slope, without any control at all. This phenomenon is related to the concept of embodiment, which originated in robotics. Embodiment refers to the idea that, when controlling a robot, one must consider not only the controller itself but the entire interacting system of controller, body, and environment. In other words, passive dynamic walkers demonstrate that, by appropriately designing the body and the environment, all of the control required for bipedal walking can effectively be outsourced from the controller to the body and environment.
Although passive dynamic walkers can be regarded as a kind of minimal example for explaining embodiment, it has also been shown that the dynamics of body–environment interaction in such systems possess extremely rich structure from the perspective of nonlinear dynamics. For example, even if the robot’s body is kept fixed, simply changing the slope angle can give rise to periodic gaits of all kinds of periods, as well as highly complex and difficult-to-predict behaviors such as chaos. Moreover, for certain slope angles, a fractal (self-similar) structure appears in which an arbitrarily small change in the initial posture can switch the eventual outcome between continued walking and falling over. In such a situation, no matter how precisely the initial posture is controlled, the subsequent behavior cannot be predicted; bipedal walking effectively behaves like throwing dice or flipping a coin. My research analyzes these dynamical properties of walking from both numerical and theoretical perspectives.
Related Publications
Related Publications
K. Okamoto, N. Akashi, I. Obayashi, K. Nakajima, H. Kokubu, K. Senda, K. Tsuchiya, S. Aoi, Sharp changes in fractal basin of attraction in passive dynamic walking, Nonlinear Dynamics, 2023.
S. Hirai, H. Mochiyama, K. Nakajima, N. Akashi, Modeling and Control of Continuum Body. In The Science of Soft Robots: Design, Materials and Information Processing (Springer Singapore), pp. 279-318, 2023.
N. Akashi, K. Nakajima, Y. Kuniyoshi, Unpredictable as a Dice: Analyzing Riddled Basin Structures in Passive Dynamic Walker, 2019 International Symposium on Micro-NanoMechatronics and Human Science (MHS), pp. 119-123 (4 pages), 2019.
3. Dynamical Analysis of Spintronics Devices and Neuromorphic Computing
3. Dynamical Analysis of Spintronics Devices and Neuromorphic Computing
- Related fields: Nonlinear Physics, Machine Learning
- Keywords: Chaos, Bifurcation Structure, Physical Reservoir Computing, Neuromorphic Computing, Spintronics
As the field of machine learning continues to advance, the computational demands placed on computers are increasing year by year. In other words, the dramatic and rapid progress of machine learning as software calls for equally fundamental advances in computing devices as its hardware counterpart. In recent years, artificial intelligence has achieved performance surpassing that of humans in a wide range of areas, including cognitive tasks such as image and speech processing as well as games. However, in terms of time and energy efficiency, humans can still be regarded as far superior. For this reason, neuromorphic computing, which imitates the neural networks of the brain, has been actively studied and developed as a next-generation computing paradigm.
Spintronics devices, which exploit the dynamics of electron spin in electromagnetic devices, have attracted attention as promising candidates for neuromorphic computing because of their advantages such as high speed, compact size, and high energy efficiency. Among them, a class of spintronics devices known as spin-torque oscillators is known to exhibit highly complex dynamics, including high-frequency oscillation, switching, and chaos. It has been shown that the dynamics of spin-torque oscillators can be exploited as computational resources for machine learning through an approach called physical reservoir computing. Our research aims to analyze the dynamics of spin-torque oscillators through numerical simulations and the theory of nonlinear dynamics, in order to explore their relationship to information-processing capability in neuromorphic computing and to identify more effective ways of utilizing them.
Related Publications
Related Publications
N. Akashi, Y. Kuniyoshi, S. Tsunegi, T. Taniguchi, M. Nishida, R. Sakurai, Y. Wakao, K. Kawashima, K. Nakajima, A Coupled Spintronics Neuromorphic Approach for High-Performance Reservoir Computing, Advanced Intelligent Systems 4: 2200123, 2022.
N. Akashi, T. Yamaguchi, S. Tsunegi, T. Taniguchi, M. Nishida, R. Sakurai, Y. Wakao, K. Nakajima, Input-driven bifurcations and information processing capacity in spintronics reservoirs, Physical Review Research 2: 043303, 2020.
T. Yamaguchi, N. Akashi, K. Nakajima, H. Kubota, S. Tsunegi, T. Taniguchi, Step-like dependence of memory function on pulse width in spintronics reservoir computing, Scientific Reports 10: 19536, 2020.
T. Yamaguchi, N. Akashi, S. Tsunegi, H. Kubota, K. Nakajima, T. Taniguchi, Periodic structure of memory function in spintronics reservoir with feedback current, Physical Review Research 2: 023389, 2020.
T. Yamaguchi, N. Akashi, K. Nakajima, S. Tsunegi, H. Kubota, T. Taniguchi, Synchronization and chaos in a spin torque oscillator with a perpendicularly magnetized free layer, Physical Review B 100: 224422, 2019.
T. Taniguchi, N. Akashi, H. Notsu, M. Kimura, H. Tsukahara, K. Nakajima, Chaos in nanomagnet via feedback current, Physical Review B 100: 174425, 2019.
4. Computation through the Bodies of Soft Robots
4. Computation through the Bodies of Soft Robots
- Related fields: Embodied Cognitive Science, Soft Robotics, Machine Learning
- Keywords: Physical Reservoir Computing, Pneumatic Artificial Muscles
Most modern robots are primarily made of rigid materials. However, for applications such as wearable devices, which require flexibility, safety, and adaptability to the environment, or rescue robots operating on uneven terrain such as disaster sites, soft materials may be more suitable. The research field that studies such soft robots is known as soft robotics, a relatively new area within robotics. A representative example of an actuator used in soft robots is the McKibben-type pneumatic artificial muscle. This actuator is a tube-shaped component made of rubber and braided mesh, and by applying air pressure inside the tube, it can produce extension, contraction, and bending motions. It offers several advantages, including flexibility, high output, and low manufacturing cost.
At the same time, soft robots also present unique challenges precisely because of their softness. For example, the nonlinear and high-dimensional bodies that enable softness make their motion difficult to predict, requiring control methods that differ from those used for conventional rigid robots. In our research, we regard the complex dynamics generated by this softness as a computational resource and propose using them for physical reservoir computing. By exploiting the intrinsic nonlinearity and hysteresis of pneumatic artificial muscles, we have achieved prediction of their own motion and pattern control. This makes it possible to outsource sensing and control computations that would otherwise be performed by an external computer to the body of the artificial muscle itself.
関連成果
関連成果
N. Akashi, Y. Kuniyoshi, T. Jo, M. Nishida, R. Sakurai, Y. Wakao, K. Nakajima, Embedding bifurcations into pneumatic artificial muscle, Advanced Science, 2024.
W. Sun, N. Akashi, Y. Kuniyoshi, K. Nakajima, Physics-informed recurrent neural networks for soft pnuematic actuators, IEEE Robotics and Automation Letters, 7 (3) pp. 6862-6869, 2022.
W. Sun, N. Akashi, Y. Kuniyoshi, K. Nakajima, Self-organization of physics-informed mechanisms in recurrent neural networks: a case study in pneumatic artificial muscles, Proceedings of 2022 IEEE 5th International Conference on Soft Robotics (RoboSoft), pp. 409-415, 2022.
W. Sun, N. Akashi, Y. Kuniyoshi, K. Nakajima, Physics-informed reservoir computing with autonomously switching readouts: a case study in pneumatic artificial muscle, Proceedings of 2021 International Symposium on Micro-NanoMechatronics and Human Science (MHS), pp. 1-6 (6 pages), 2021 (Best Paper Awards).
R. Sakurai, M. Nishida, H. Sakurai, Y. Wakao, N. Akashi, Y. Kuniyoshi, Y. Minami, K. Nakajima, Emulating a sensor using soft material dynamics: A reservoir computing approach to pneumatic artificial muscle, in Proceedings of 2020 3rd IEEE International Conference on Soft Robotics (RoboSoft), pp. 710-717, 2020.
Page updated
Google Sites
Report abuse