Quadcopter Soft Vertical Landing Control with Hybrid Physics-informed Machine Learning

摘要

The design of classical controllers for quadcopters usually takes advantage of their easily linearized rigid body dynamics. Unfortunately, in certain parts of the flight, the interaction between the quadcopter and the environment is difficult to model. In the case of landing, the ground effect exerts a lift contribution that can degrade the performance of classical controllers for soft landing applications. In this paper, we combined a classical proportional-integral-derivative control implementation together with a machine learning model that compensates for the lift contribution of the ground effect. We first characterized the curves for the quadcopter propulsion system using flight data out of ground effect. Then, we used low altitude flight data to build a neural network model for the lift contribution due to ground effect. We implemented our hybrid controller using TensorFlow and deployed it on a NVIDIA Jetson TX2, which runs as an integrated component of the onboard controller. The results shows a significant improvement when compared to the classical controller with regards to the ability to manage the interactions between the quadcopter and the ground in soft landing.