弱引力环境四足机器人的强化学习姿态机动控制  

Reinforcement Learning Control for Quadruped Robots Attitude Maneuver under Weak Gravitational Environment

作　　者：高海波[1] 齐骥 苏桓立 霍明英[2] 于海涛[1] 邓宗全[1] GAO Haibo;QI Ji;SU Huanli;HUO Mingying;YU Haitao;DENG Zongquan(State Key Laboratory of Robotics and System,Harbin Institute of Technology,Harbin 150080,China;School of Astronautics,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学机器人技术与系统国家重点实验室,哈尔滨150080 [2]哈尔滨工业大学航天学院,哈尔滨150001

出　　处：《宇航学报》2024年第10期1676-1684,共9页Journal of Astronautics

基　　金：国家自然科学基金(52175011,U22B2013,12272104);中国航天科技集团有限公司第八研究院产学研合作基金(SAST2022-014);中央高校基本科研业务费专项资金资助(HIT.DZJJ.2023071)。

摘　　要：针对四足机器人在小天体表面跳跃巡游探测过程中的空中姿态控制问题,提出了一种基于强化学习的无模型控制方案。在小天体的弱引力环境下,机器人单次跳跃的滞空时间很长,机器人需要利用这段时间修正起跳带来的姿态偏差以实现安全着陆,或改变偏航角以改变未来行进方向。提出的基于近端策略优化(PPO)算法训练的控制器以猫落反射为启发,无需引入额外的姿态控制执行器。仿真结果表明,训练得到的控制器可通过控制四足机器人摆动腿部关节实现三轴姿态机动。设计并搭建了四足机器人气浮微重力模拟试验平台,通过样机试验,控制策略模型的有效性得到了验证。A model-free control method based on reinforcement learning is proposed to address the attitude control problem of quadruped robots during hopping and exploration on small celestial bodies.In the low-gravity environment of these bodies,the robots undergo extended periods of free fall during each jump.It is essential to utilize this time to correct any attitude deviations induced by the jump,ensuring a safe landing or adjusting the yaw angle to modify their future trajectory.The proposed controller,inspired by the cat righting reflex,is trained using the Proximal Policy Optimization(PPO)algorithm and does not necessitate additional attitude control actuators.Simulation results demonstrate that the trained controller can achieve three-axis attitude maneuvering by controlling the swinging leg joints of the quadruped robot.Furthermore,a microgravity simulation test platform for quadrupedal robots employing air suspension technology has been designed and constructed,validating the effectiveness of this control strategy through prototype experiments.

关 键 词：星球探测机器人 小天体探测 强化学习 微低重力模拟 

分 类 号：V47[航空宇航科学与技术—飞行器设计]