文章基本信息

标题：Actuation Strategy of a Virtual Skydiver Derived by Reinforcement Learning
本地全文：下载
作者：Anna Clarke ; Per-Olof Gutman
期刊名称：IFAC PapersOnLine
印刷版ISSN：2405-8963
出版年度：2020
卷号：53
期号：2
页码：1569-1574
DOI：10.1016/j.ifacol.2020.12.2187
语种：English
出版社：Elsevier
摘要：AbstractAn innovative approach of training motor skills involved in human body flight is proposed. Body flight is the art of maneuvering during the free fall stage of skydiving. The key idea is gradually constructing the movement patterns which are the combinations of body degrees-of-freedom that are activated synchronously and proportionally as a single unit, and turning this process into a coaching strategy. The proposed method is iterative: at each skill level an optimal movement pattern is constructed from the basic elements of the current movement repertoire. The free-fall maneuvers of each learning stage can be executed using any one of the basic elements. The construction has two stages: 1. tracking the desired maneuver while the body is actuated by each one of the basic patterns; 2. finding an optimal combination of these patterns to form a new way of body actuation. This hierarchical design resolves stage 2 by Reinforcement Learning with pure exploration and a minimal number of episodes. The method was tested in a Skydiver Simulator and resulted in deriving a movement pattern that showed a superior performance of the studied maneuver. The states and the reward of the Reinforcement Learning algorithm were converted into motor learning aids.
关键词：Keywordsautonomous systemssimulatorsmultiple degrees-of-freedomreinforcement learning