Reinforcement learning control for the swimming motions of a beaver-like, single-legged robot based on biological inspiration
Reinforcement learning control for the swimming motions of a beaver-like, single-legged robot based on biological inspiration
复制标题
基于生物启发的类海狸单足机器人游泳动作的强化学习控制
DOI：
10.1016/j.robot.2022.104116
复制
发表时间：
2022-05-07
影响因子：
4.3
通讯作者：
Hu, Huosheng
中科院分区：
文献类型：
--
作者：
Chen, Gang;Lu, Yuwang;Hu, Huosheng
复杂的流体动力学建模和分析被认为是水下仿生机器人运动研究的绊脚石。近年来，强化学习技术已应用于未知环境下的机器人运动控制。然而，机器人在学习过程中可能会以非常规或危险的方式行事。这些行为增加了训练难度，降低了训练效率。在这项研究中，提出了一种仿生强化学习控制方法。它通过海狸的离散游泳运动实现了机器人自学习的运动策略，而不需要建立水下机器人的流体动力学等运动模型。仿生模型进一步减少了机器人在强化学习过程中的无效动作，提高了训练效率。实验结果验证了所提出的机器人平台的环境适应和自学习能力，证明了基于生物启发的机器人游泳强化学习控制方法的有效性。该研究成果为水下仿生机器人的运动控制提供了新思路，进一步推动人工智能在水下机器人中的应用。 (c) 2022 Elsevier B.V. 保留所有权利。
Complex hydrodynamic modeling and analysis are considered as stumbling blocks in the motion study of underwater bionic robots. In recent years, reinforcement learning techniques have been applied for robot motion control in unknown environments. However, robots may act in an unconventional or dangerous manner during the learning process. These actions increase the training difficulty and decrease the training efficiency. In this study, a biological-inspired reinforcement learning control method is proposed. It realizes the self-learning movement policy of the robot with discretized swimming motions of a beaver without the need to establish motion models, such as hydrodynamics, of underwater robots. The biological-inspired model further reduces the robot's ineffective movements during the reinforcement learning and improves training efficiency. The experiment results verify the environmental adaptation and self-learning ability of the proposed robot platform and proves the effectiveness of the reinforcement learning control method for robotic swimming based on biological inspiration. This study's findings provide new ideas for the motion control of underwater bionic robots and further promote the application of artificial intelligence in underwater robots. (c) 2022 Elsevier B.V. All rights reserved.