Active-Matter Physics for Uncovering Principles of Living Systems: From Fundamental Understanding to Functional Implementation (Part II)
Reservoir Computing for Molecular Swarms
2026 Volume 66 Issue 2 Pages 113-114
Details
2026 Volume 66 Issue 2 Pages 113-114
リザーバーコンピューティングは，複雑な動的システムを利用して計算する方法の一種である．特に分子プログラミングに応用でき，分子ロボットの群れの制御にも活用できる．ここでいう群れは，要素となる個体（ロボット）の集団であり，局所的な相互作用により自己組織化を行う．
Reservoir computing is a Machine Learning approach to deal with the complexity of training Recurrent Neural Networks, highly non-linear mathematical models, but has been since then applied to a large range of complex dynamic systems1). The main strategy is to fix the central part of the system (the reservoir) and only train the output layer (called the readout). As long as it implements a reservoir with relevant properties, such as memory capacity, any physical system is a potential target for this computing approach, providing a flexible programming strategy. As previous work shows, molecular computing systems are among those promising approaches1),2), allowing us to create specialized computing systems at the molecular level, with potential applications to the control of large molecular swarms.
The development and control of large robotic swarms is a current robotics grand challenge3). Reducing the size and complexity of individual units drastically reduces their price and energy consumption compared to a monolithic robot, at the cost of limited hardware capabilities. Those limits, however, are overcome by the emergence of collective behaviors of the swarm itself, mimicking the natural world. As the swarm grows larger, its intrinsic computational capabilities increase in theory4), but that theoretical insight requires to reach million units or more, a number that is only realistic by going to the molecular scale.
That approach, which takes advantage in the recent advances in molecular robotics, has been demonstrated through colloids, microbeads5),6), microtubules7),8), and other molecular systems (see ref. 9 for instance).
In this context, structures like microbeads or microtubule form the main “body” of the robot, functionalized with other molecules (typically DNA5)-8)) acting as sensors and actuators. Chemical reactions, such as hybridization (two complementary DNA strands attaching to each other) or denaturation (the reverse reaction), allow those molecules to change configuration based on external inputs or interactions with other robots. The control of the robot then takes place through Chemical Reaction Networks (CRNs), providing a way to cascade information from the sensors through processing units into a modification of the actuators.
We are then left with two main challenges: 1) implementing input/output strategies that allow us to interact directly with the system while it remains closed, and 2) providing approaches to design controllers for specific tasks, considering the highly non-linear and dynamic nature of CRNs. This topic paper will describe recent efforts to apply Reservoir computing to solve both challenges
One of the main issues when attempting to use a robotic swarm for a specific application lies in the ways to interact with the system as a whole. While leaving the system completely closed, relying only on initial conditions to perform its task is an option, it may be unsatisfying for cases where external conditions would lead the user to dynamically alter behavior. That issue is even more pressing in the case of molecular robots, as typical broadcast strategies used with their electronic counterparts are unavailable.
In previous work, one approach has been to rely on azobenzene modification to the backbone of DNA molecules used in the robot controllers7),9). When irradiated with UV light, the azobenzene molecules switch to a cis configuration, thus preventing the hybridization of DNA strands, thus cutting reaction paths in the controller. This effect typically leads to the destabilization of the swarm, the drop of cargo, and so on. That configuration change can be reversed through exposure to visible light, overall providing a single binary input to the system. While additional information may be conveyed through the temporality of the input (thus encoding a binary string instead), direct application may be challenging8).
Another type of input currently considered is temperature, as it provides a more flexible way to interact with the dynamics of the controller. On top of controlling the stability of double-stranded structure, as is possible with azobenzene, the paths corresponding to different reactions may be independently altered when using CRNs built from enzymatic reactions, yielding various dynamics2),10). That approach may be used to encode binary strings like the previous strategy10). Moreover, while the impact of temperature on the system is highly non-linear, making rational design difficult, the approach fits well reservoir computing2).
Alternatively, one approach may be to rely on microfluidics, thus keeping the system open. While that approach may not be fit for all applications, it provides the most flexibility, allowing the use of DNA strands (the signals of the controller) directly as input and output.
The second challenge is to provide enough computing power to the controller of the swarm, allowing the implementation of tasks such as targeted self-aggregation5),6) or cargo transport7) and sorting8).
The DNA strands required to implement the controller can either be localized (e.g., attached to a microbead or microtubule) or left to diffuse freely in the environment. In traditional swarm robotics, however, the controller is supposed to be part of the robot (a concept named embodiment). As such, we will focus on the former approach.
Previous attempts showed that DNA-functionalized microbeads did provide promising computing capabilities. However, for more complex controllers, the surface required to fit the whole system prevents the beads from moving on their own. Meanwhile, DNA-functionalized microtubules have proved great candidates for implementing molecular swarms, but lack the surface required for functionalization too. Our approach has thus been to combine the best of both worlds: functionalized beads provide the computational part while being carried by microtubules (Fig. 1). That approach offers additional flexibility in allowing a group to “hot swap” its controller by dropping a given bead and picking up another. Finally, our recent results show that the swarming behavior itself can be used as the output of a reservoir, thus allowing us to apply the reservoir computing approach to the control of the overall swarm.
Recent advances have shown that the implementation of massive swarms at the molecular scale is possible, thus opening the door to the in-vitro implementation of so far theoretical systems from swarm robotics. While interacting with and overall controlling such swarms is challenging due to the widely different programming paradigms, reservoir computing offers a promising approach to tackle those issues.
AUBERT-KATO Nathanael（オベル加藤 ナタナエル）
Department of Information Sciences, Ochanomizu University
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/66-2ie9p01.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
