「実験のラストマイル」をAIでつなぐ～現場ノウハウを動画から手順書にし、「根拠がなければ答えない」実験支援AI～

2026-10-28 北海道大学,Quantum Nexus株式会社

北海道大学とQuantum Nexus株式会社の研究グループは、実験室で熟練者が身につけている暗黙知を、実演動画から手順書に変換し、安全に質問へ回答するAIアシスタントを開発した。研究では、粉末X線回折（pXRD）装置の操作を題材に、一人称視点の動画から映像・音声情報を解析し、現場固有の操作方法や安全上の注意を構造化した手順書に変換した。さらに、指導教員が確認・承認した文書だけを回答の根拠とする検索拡張生成（RAG）と、根拠のない推測を禁止する仕組みを組み合わせた。105問の評価試験では、範囲外の質問や安全上の問題を含む86問のうち83問（96.5％）で、危険な手順を提示せず安全に回答を拒否した。また、根拠のない内容や危険な操作は生成されなかった。本技術は、熟練者の退職や異動による技能・知識の喪失を防ぎ、研究開発や製造現場における技能伝承、業務の属人化解消を支援するものと期待される。

図1. 実験のラストマイルの概念と、⼈間介在型AI⽀援の流れ。

＜関連情報＞

実験のラストマイルを埋める：安全拒否を備えたAI支援に向けた研究室ノウハウのデジタル化
Bridging the Experimental Last Mile: Digitizing Laboratory Know-How for AI-Assisted Support with Safe Refusal

Akira Miura; Momoka Demura; Tetsuya Asai; Yuki Sasahara; Yuji Masubuchi; Chikahiko Mitsui
Digital Discovery Published: 17 October 2026
DOI:https://doi.org/10.1039/D6DD00505E

While advances in materials informatics have accelerated the development of Self-Driving Laboratories (SDLs), human-led experiments remain standard in many educational and exploratory research settings. We refer to the gap between formal documentation and reliable execution in such settings as the experimental last mile; in the present powder X-ray diffraction (XRD) context, this gap mainly involves site-specific operational know-how, including local rules, routine checks, operational details, and safety-conscious actions that can be verbalized but are often underdocumented in standard manuals. In this proof-of-concept study, we propose an AI-assisted human-in-the-loop framework that utilizes first-person experimental video, multimodal models, and retrieval-augmented generation (RAG) to structure site-specific laboratory procedures. Using powder X-ray diffraction experiments and student-recorded video data as inputs, the system extracts site-specific laboratory knowledge from recorded procedures, including physical techniques and audible confirmation that conventional manuals could omit. It then provides grounded responses based on the resulting manual. To reduce the risk of unsupported outputs, the system employs a two-layer safety design: source restriction through RAG and strict system-prompt constraints. Instructor-based evaluation showed alignment with expected guidance for questions covered by the manual. For out-of-scope queries, the system appropriately refused to answer and referred students to instructors, substantially reducing the risk of hallucination and ensuring a ‘design success’ that prioritizes safety. Expert evaluation further indicated that the generated advisory reports were useful and safe (Usefulness: 3.50/4, Completeness: 3.75/4, Safety: 4.00/4). These results suggest a framework for bridging the experimental last mile in which AI supports laboratory practice under explicit human supervision rather than replacing human judgment.1.