Shinji Ito

Mathematical Foundations of Online Learning Technologies Adaptable to Real-World Environments

Grant No.：JPMJPR2511

Researcher

Shinji Ito

Outline

In this study, we aim to develop machine learning methods that can autonomously adapt and achieve high performance in online learning settings—including bandit algorithms and online reinforcement learning—even when the assumptions about the operating environment are unclear. Our goal is to establish a theoretical and computational foundation for online learning that reconciles rigorous performance guarantees with practical efficiency, without relying on the traditional distinction between stochastic and adversarial environments. The proposed framework will also be capable of handling complex, high-dimensional decision-making problems as well as intermediate or non-stationary environments.

Yusheng Wang

Development of Acoustic-Sensing Intelligent Agents in Fluid Environments

Grant No.：JPMJPR2512

Researcher

Yusheng Wang

Outline

This research aims to model the influence of fluid dynamics on both acoustic sensing and robotic dynamics, and to construct a next-generation underwater robot simulator. By exploiting acoustic information, the simulator enables non-contact perception of surrounding fluid states and integrates this capability as a foundation for advanced robot control. Within this environment, data-driven dynamic models can be learned efficiently and combined with model predictive control, allowing robots to adapt robustly to complex and unpredictable underwater conditions. Through these developments, the research seeks to bridge the gap between sensing, simulation, and control, ultimately contributing to the realization of autonomous underwater robots that can achieve stable and reliable operation even under strong fluid disturbances in real-world coastal environments.

Keisuke Okumura

Large-Scale Multi-Robot Coordination via Neuro-Guided Multi-Agent Search

Grant No.：JPMJPR2513

Researcher

Keisuke Okumura

Outline

The aim of this project is to develop methodologies for the automatic design of coordinated behaviour in multi-robot systems through the integration of search and learning schemes. Specifically, we intend to combine imitation learning with recent advances in scalable multi-agent search algorithms to iteratively refine agent-wise policies. To demonstrate the feasibility of this approach, the developed techniques will be deployed with a hundred-scale robot swarm to solve coordinated multi-robot navigation problems.

Taisuke Kobayashi

Unified Regulation of Dual Dilemmas in Lifelong Active Learning

Grant No.：JPMJPR2514

Researcher

Taisuke Kobayashi

Outline

I aim to realize robotic systems that autonomously continues learning in the real world by integrating lifelong learning and active learning, which have often been discussed independently. Specifically, by focusing on the duality inherent in the respective dilemmas faced by lifelong learning and active learning, I develop a general-purpose algorithm enabling the systems to act and learn with a desired balance that does not contradict each other’s dilemmas.

Ryo Takahashi

Motion-durable, full-body networking for distributed skin intellgience

Grant No.：JPMJPR2515

Researcher

Ryo Takahashi

Outline

This research aims to create a “textile-based, full-body wireless power and data network” capable of operating a battery-less, wire-free, high-density skin device that can conform to the deformation of the skin during vigorous human or robotic movement. These e-skin devices are designed to conform to skin deformation during vigorous human or robotic movement and collect physiological and tactile data for AI analysis. The network will be engineered to maintain consistently high communication and power transfer performance, irrespective of garment deformation. Furthermore, it will be capable of reading data from the high-density array of skin devices and executing functions such as AI analysis locally.

Tomohiko Nakamura

Variable-Resolution Audio Processing Framework Adaptive to Acoustic Scenes

Grant No.：JPMJPR2517

Researcher

Tomohiko Nakamura

Outline

To keep audio media processing technologies running reliably in real-world environments, systems need to flexibly handle changes in connected devices and surrounding acoustic environments. Such changes occur not only when devices or systems are updated or replaced, but also when acoustic scenes can change depending on the time of day. This project aims to establish a deep learning framework for time-series signals that employs a parameter representation independent of input and output time resolution. Building on this framework, we seek to realize audio scene understanding technologies that can autonomously adapt to varying conditions by flexibly adjusting their time resolution.

Mai Nishimura

Self-Organizing Episodic Memory System for Lifelong Learning

Grant No.：JPMJPR2518

Researcher

Mai Nishimura

Outline

The Self-Organizing Episodic Memory system provides a foundational infrastructure enabling real-world AI systems to accumulate experiences and engage in lifelong learning. Unlike conventional paradigm that constructs memory (i.e., offline datasets) statically prior to training, this system implements an integrated memory process comprising experience accumulation, self-organization, and semantic recall. These processes are autonomously orchestrated through generative search and task-driven optimization over memory space, enabling situation-adaptive skill acquisition and continuous AI evolution. This fundamentally transforms how AI systems manage and leverage episodic memory throughout their operational lifetime.

Tatsuya Haga

Modeling and application of internal thought processing in the brain

Grant No.：JPMJPR2519

Researcher

Tatsuya Haga

Outline

Recently, Chain-of-Thought (CoT) has enabled large-scale language models to perform inferences through complex thought processes. In this research, we build a theory of brain-like CoT by modeling the generation of internal thought processes in the hippocampus-prefrontal network. Furthermore, by implementing mind wandering that is characteristic of the brain, we aim to build a machine learning model that can learn from small amounts of data and think flexibly like the human brain.

Daisuke Hisano

Task-Oriented AI-Integrated Communication System for Extreme Communication Environments

Grant No.：JPMJPR251A

Researcher

Daisuke Hisano

Outline

This study aims to establish an AI-integrated communication system platform that enables the stable and efficient acquisition and transmission of multimodal data in extremely resource-constrained real-world environments. In particular, it focuses on three types of extreme environments: underwater communication, deep-space communication, and narrowband cellular communication, and seeks to realize a low-latency, highly reliable communication infrastructure capable of transmitting information obtained from these environments in a format that can be effectively utilized by AI systems.