Research On Quadrotor Control Based On Deep Reinforcement Learning

Quadrotor is a multi-rotor UAV that is lifted and propelled by four rotors.Because of its vertical take-off and landing,simple mechanical structure,flexible control and other characteristics,quadrotors are widely used into various fields.However,because quadrotor is a highly complex system,the design of controller with superior performance is the key to realize the control of quadrotor.In this thesis,the design and implementation of quadrotor controller was completed based on the traditional PID algorithm and the DRL algorithm.Then,we compared and analyzed the control performance of each controller.Firstly,the dynamic model of quadrotor was established according to the principle of flight.On this basis,we designed the controller of quadrotor by using PID control algorithm.Then,the parameter of PID was set by engineering trial method,so that the dynamic and static performance of the system can meet the control requirements.The premise of designing quadrotor controller based on traditional control methods is that it is necessary to fully understand the dynamic characteristics of quadrotor to establish the complex model of quadrotor.To resolve this issue,we introduced the DRL algorithm into quadrotor control to realize the design of quadrotor controller.The algorithm is an end-to-end learning algorithm that combines DL(Deep Learning)with RL(Reinforce Learning)to realize from perception to action.The algorithm only needs to get the prior knowledge about the model to determine the relationship between the input and output of the system,without having to consider the specific model of the system.After that,the structure of neural networks,learning rate and other hyperparameters were optimized to realize the control of quadrotor by crossvalidation.The simulation results verify the control performance of the quadrotor controller based on DRL algorithm.Finally,this thesis designed a set of experimental platform of quadrotor control based on Machine Learning.All modules were drove by the firmware.Then a communication protocol was designed to make the communication between the upper computer and quadrotor more reliable.The testing program of the upper computer was designed to realize the joint debugging of the whole system,which lays a foundation for further optimizing the algorithm and improving the control performance.