2023 Fiscal Year Final Research Report

Unified Understanding through Analysis and Verification of Human Environmental Adaptation Learning Methods

Research Project

PDF

Project/Area Number	20KK0256
Research Category	Fund for the Promotion of Joint International Research (Fostering Joint International Research (A))
Allocation Type	Multi-year Fund
Review Section	Basic Section 20020:Robotics and intelligent system-related
Research Institution	Tohoku University
Principal Investigator	Hayashibe Mitsuhiro 東北大学, 工学研究科, 教授 (40338934)
Project Period (FY)	2021 – 2023
Keywords	運動学習 / 模倣学習 / 深層強化学習 / 同期現象 / 運動シナジー / CPG / 環境適応
Outline of Final Research Achievements	The main theme of this research, "Mechanism of Expression of Diverse Movements through Environment-Adaptive Learning Independent of Environment Models and Oscillator Models," aims to elucidate the mechanism of environment-coordinated motion control by investigating how environment-adaptive rhythmic motion generation can be achieved through what computational methods. This international collaborative research has yielded results contributing to a unified understanding of human environmental adaptation learning methods from three perspectives. With Professor d'Avella from Italy, achievements have been made in understanding and replicating the synergy structure of whole-body movement expression processes. With Professor Burdet from the UK, achievements have been made in enhancing motor learning efficiency through shared tactile information. With Professor Ijspeert from Switzerland, foundational technologies for developing AI capable of mimicking biological movements have been developed.
Free Research Field	ロボティクス
Academic Significance and Societal Importance of the Research Achievements	近年、深層強化学習や模倣学習をそれぞれ用いたロボット制御の応用研究が活発に行われ注目されています。深層強化学習を活用する場合には環境適応可能な運動が生成できるものの、広大な入力空間の探索に膨大な計算コストを要することが問題となります。一方、模倣学習を用いる場合には学習した運動に近い範囲に環境適応性が制限されるという問題が一般的に知られています。今回の提案手法は深層強化学習と模倣学習の両面の利点を生かすことができ、またその欠点を補いあうことができる新しい運動生成の手法となり、多自由度系で生体の自己組織的な振る舞いの生成をAIにより実装する技術につながります。