P-01
Introduction to Autonomous Lab: Autonomous Experimental System
Taichi Tomono (SHIMADZU CORPORATION)
This is an overview introduction of Autonomous Lab, an autonomous experimental system that aims to automatically run the Design→Build→Test→Learn experimental cycle at Shimadzu Corporation. We will introduce the contents of the prototype targeting biomanufacturing as an initial target.
P-02
Automated Exploration of Novel Quantum Chemistry Methods Using Large Language Models
Masaya Hagai (TOYOTA MOTOR CORPORATION)
While quantum chemical calculations enable molecular-level simulations, the introduction of approximation methods to reduce computational costs is essential for handling large molecular systems in realistic time. However, creating useful approximation methods faces barriers such as interdisciplinary knowledge spanning physics, chemistry, mathematics, and information science, difficulty predicting execution costs, and high implementation costs. In this study, we propose a method that automates the conception, implementation, and verification of new quantum chemistry methods using large language models, enabling automated exploration of promising approximation methods by quickly trying diverse interdisciplinary ideas.
P-03
Development of Domain-Specific Foundation Models for AI-for-Science
Yuna Oikawa (The University of Tokyo, Tsuda Laboratory)
We propose two large-scale models in the AI-for-Science domain: (1) GPepT, a foundation peptide language model for peptide mimetics design incorporating non-natural amino acids, and (2) aLLoyM, a large language model specialized for phase diagram prediction of metallic materials, integrating knowledge acquisition, generation, and prediction capabilities in each scientific domain.
P-04
Development of Crystal Structure Generation Framework Based on Natural Language Instructions
Yusei Ito (The University of Osaka)
Designing crystal structures that achieve desired properties is an important challenge in materials exploration. Existing methods using generative models such as MatterGen have shown effectiveness in structure generation conditioned on numerical property values, but since they assume explicit quantitative information, they have limitations in flexibly handling complex and multifaceted design goals. In this study, we propose a text-conditioned structure generation framework that enables direct generation of crystal structures from natural language instructions.
P-05
Object-Centric World Model and Policy Learning Considering Influence Relations
Haruhiro Nishimoto (The University of Osaka)
Deep reinforcement learning agents have extremely low sample efficiency, making real-world application difficult. To address this problem, many model-based reinforcement learning methods that learn world models have been proposed, achieving some success. However, in complex environments with multiple objects, it remains challenging for reinforcement learning agents to acquire world models. In this study, we propose a reinforcement learning agent where the world model, policy function, and value function are all composed of Transformers that handle object-centric representations, confirming its effectiveness in multiple benchmarks.
P-06
Development of Tailor-Made Experiment Automation Tools for Cell Culture
Atsushi Shibai (RIKEN Special Postdoctoral Researcher)
Targeting microbial culture experiments, we develop tailor-made automation tools that combine GUI wrappers with low-cost robot transport while leveraging existing culture and measurement equipment, verifying utility and versatility in drug resistance evolution experiments and stress tolerance acquisition experiments.
P-07
Flexible Control of Conditional Generative Models
Kuniaki Saito (OMRON SINIC X Corporation)
Models that generate data such as language and compound structures are expected to obtain desired outputs from various conditions. In this poster, we present findings on conditional generative models, focusing on Vision-Language Models that generate captions from images. Note that the results in this poster have been presented at ICCV2025.
P-08
ROS2/PLC Integration for Automating Existing Experimental Equipment and Construction of Automated Arc Furnace
Kensei Terashima (NIMS)
Many existing experimental devices still have technical barriers to plug-and-play automation. We built a framework using ROS2 and PLC, where LAN-communicable devices are handled via ROS2, and those that cannot are handled via PLC as ROS2 nodes, enabling experimental equipment to operate independently and in parallel, coordinating as needed. We created an automated solid bulk material synthesis system using an arc furnace.
P-09
Discovering New Theorems via LLMs with In-Context Proof Learning in Lean
Kazumi Kasaura (Research Engineer, OMRON SINIC X Corporation)
This study focuses on LLMs' ability to discover new theorems. We propose a "conjecture-proof loop" pipeline that automatically generates mathematical conjectures and proves them in Lean 4 format, a type of theorem proving assistant. The key feature of this method is that it generates and proves further conjectures using context that includes previously generated theorems and their proofs. This enables generation of more difficult proofs through in-context learning of proof strategies without changing the LLM's parameters.
P-10
Modeling Coupled Systems and Multi-Physics Using Deep Learning
Khosrobian Lazmik Arman (The University of Osaka, Graduate School of Engineering Science, Master's 2nd year)
Deep learning has achieved great success in modeling dynamical systems with unknown governing equations. However, application of existing methods has been mainly limited to mechanical systems, and they cannot directly model coupled systems because they treat systems as a whole. We propose a deep learning model using Dirac structures that can uniformly represent systems spanning multiple domains such as mechanical systems and electrical circuits, as well as the interactions and constraints of each element. Experimental results show that the proposed method appropriately learns interactions between elements and has better long-term prediction performance than existing methods.
P-11
Development of Next-Generation Experimental Robot for Distributed Development
Takaaki Horinouchi (Engineer, RIKEN Center for Biosystems Dynamics Research)
Toward the development of AI robot-driven science, we aim to provide robot labs and system infrastructure that enable learning and generation of diverse scientific research data as technology that autonomously and actively learns. To this end, we are developing Ardea, a next-generation experimental robot system compatible with distributed development of life science experiments, and LabCode, an experiment description language. In the life science field, diverse experimental protocols exist, and distributed development and resulting network effects are considered important elements for efficient implementation. We will introduce the development status toward its demonstration.
P-12
A Minimal Open-Access Platform for Laboratory Automation Training
Daiki Noguchi (Visiting Researcher, RIKEN Takahashi Laboratory)
In this study, we developed a low-cost, highly reproducible laboratory automation (LA) education system and created DIY-LA equipment based on open source that links the OT-2 liquid handler and Absorbance-96 plate reader. This system enables easy learning of the basic elements of LA - "sample preparation," "measurement," and "experimental design" - both individually and as an integrated experimental system. High educational effectiveness was confirmed in Bayesian optimization training for undergraduate students at Keio University.
P-13
NIMO: Software Supporting Automated Autonomous Experiments
Shoji Yamaguchi (NIMS Center for Energy and Environmental Materials)
We developed NIMO, a general-purpose software to easily form closed loops for automated autonomous experiments, and released it as open-source software. In NIMO, robot experimental equipment and exploration AI are treated as independent modules, designed to enable various automated autonomous experiments by flexibly selecting combinations of robot experimental equipment and AI algorithms. In particular, multiple options for exploration AI algorithms are implemented, allowing various types of exploration to be conducted with a single robot experimental device.
P-14
Introduction to JST PRESTO "Research and Development Process Innovation"
Ichiro Takeuchi (Professor, Nagoya University)
This poster introduces the JST PRESTO area "Infrastructure Construction and Practical Application for R&D Process Innovation through AI/Robots (Research and Development Process Innovation)." (Display only)