Water-surface take-off control of aquatic-aerial trans-domain robot with reinforcement learning [基于深度强化学习的水空跨域机器人水面起飞运动控制]

作者：Huo Y.-J.; Li Y.-P.; Feng X.-S.

来源：Electric Machines and Control, 2021, 25(12): 139-150.

DOI：10.15938/j.emc.2021.12.015

• Aquatic-aerial trans-domain robot
• Aquatic-aerial trans-domain transition
• Deep reinforcement learning
• Gaussian Process
• Motion control
• Quad tiltrotor robot

The aquatic-aerial trans-domain robot's transition process of water-to-air(water surface take-off) poses a new challenge to the control due to the influence of propeller thrust attenuation, various buoyancy, and fluid dynamics between different water and air in the near-surface environment. In this paper, a robot surface buoyancy estimation, a model of water surface propeller thrust and a surface take-off method based on deep reinforcement learning(DRL) are proposed. On this basis, to solve the problem of inabilities with propellers' thrust attenuation, a method of water surface take-off in stages were proposed. The effectiveness of the phased surface take-off method based on DRL was demonstrated in simulation. The experiment shows that with DRL control, direct water surface take-off without thrust attenuation and water surface take-off in stages with thrust attenuation confirm effectiveness, and the robot achieves water surface take off with position at 2 m in high, roll and pitch less than 5°(without thrust attenuation), roll less 5° and 65° pitch(with thrust attenuation).

• 单位
  机器人学国家重点实验室; 中国科学院大学