Case Studies
Verification of UAV control algorithm and collection of data set
Unmanned System Laboratory, China University of Petroleum
4m*4m*2m
Flight control, algorithm verification, reinforcement learning
Rotor UAV
8 Mars 2H motion capture cameras

With the acceleration of the opening of low-altitude airspace in China and the development of sensors, automatic control, computers and other technical fields, micro-unmanned aerial vehicles have developed rapidly in recent years. As an important member of UAV family, because of its unique flight characteristics, such as hovering in the air, taking off and landing vertically, flying on the ground, it has become the focus of domestic and foreign universities and research institutions.

1629699997775805.png

A complete control system needs controlled objects, sensors and controllers to form a complete closed loop. Corresponding to the control system of quadrotor UAV, it is necessary to have UAV that can receive wireless communication, sensors that can get UAV status information (attitude information and position information of UAV), and ground control stations that send commands. In outdoor environment, the positioning of UAV is usually realized by GPS. However, GPS can't be used in the laboratory environment of algorithm verification stage, and the requirement of positioning accuracy in small laboratory environment is much higher than that in outdoor environment, so it is very important to find a high-precision position and pose feedback solution of control loop.

1629700011588343.png

Therefore, the Unmanned System Laboratory of China University of Petroleum adopted NOKOV optical 3D motion capture system to locate the UAV. NOKOV metric motion capture system locates Marker points placed on unmanned aerial vehicles through a moving camera emitting infrared light with a certain wavelength, and the accuracy of outputting marker position information can reach sub-millimeter level. However, the distance between Markers on UAV is fixed, so UAV can be constructed as a rigid body by multiple markers, and the dynamic capture system can get the six degrees of freedom information of rigid body in real time-roll angle, pitch angle and yaw angle. By sending the pose information to the ground station, the control algorithm of the ground station can calculate the motion parameters, estimate the waypoint position, issue the waypoint instruction, and the UAV will continue to move after receiving the instruction, and the data information at the next moment will be captured and output to the ground station, thus realizing the indoor flight control of the UAV.

After the closed-loop control is formed, an autonomous control experiment can be designed to verify the feasibility and stability of the algorithm. For example, in an autonomous hovering experiment, the hovering position coordinates of UAV can be set as fixed values, and the roll/pitch/yaw angles are all 0. After many experiments, the position deviation and attitude angle data of UAV in autonomous hovering state are measured to verify the control algorithm.

If the stability and control accuracy of the algorithm can not meet the requirements, the reinforcement learning of the algorithm can be performed. In the pre-training stage, the NOKOV metric motion capture system is used to collect pose information in real environment as the training data set of intelligent algorithm, and the reward value is constructed with specific pose data until the depth model converges after a period of training, thus achieving a more stable control effect.


Contact Us

  • For more information, please use the form here.

  • Application *
  • Submit
Contact us

Contact us