一种通过强化学习的四旋翼姿态控制算法

Quadrotor Attitude Control Algorithm Based on Reinforcement Learning

本文提出通过深度强化学习实现四旋翼姿态控制.通过对深度强化学习中的近端策略优化实现姿态稳定的控制学习任务.飞行器直接通过神经网络进行训练,直接将状态输入映射到控制输出.其次提出一种崭新的算法通过将传统的控制模型引入来提高强化学习算法的训练速度.经过实验验证,该算法可以有效控制四旋翼在任意姿态下的稳定性,且该算法可以在比一般强化学习方法更具泛用性更快收敛.

In this paper,through deep reinforcement learning,a quadrotor attitude control algorithm was proposed.Through the optimization of the proximal policy optimization in deep reinforcement learning,the learning of control task with stable attitude is realized.The aircraft is trained directly through the neural network and directly maps the state input to the actuators.Secondly,a newalgorithm is proposed to improve the training speed of the reinforcement algorithm by introducing the traditional control model.Experiments showthat the algorithm can effectively stabilize the quadcopter in any attitude,and the algorithm can be more general and faster convergence than the general reinforcement learning algorithm.