Quadrotor Auto Trimming based on Deep Reinforcement Learning and Digital Twin

In flight systems, one of the most common issues is the lack of precise placement of the robot’s sensors in the three XYZ axis at the start of the flight. Without trimming, the output variables could be changeable without any command. In this situation, the pilot must manually trim based on their try-and-error experiments at the beginning of the flight and keep adjusting them until the outputs are perfectly balanced. Nevertheless, manual trimming could always be a challenging process that takes a long time and requires more precision. Accordingly, some efforts have been put into automating it by autopilot systems recently. However, they still have problems such as being useless in the real world for general gray-box systems, real-time trimming, and unknown environments. Our goal in this paper is to find the unbalanced initial bias in one of these flight systems, quadrotors, with increased accuracy, high speed, and greater reliability. The investigation consists of applying a non-classical auto-trimming approach in real time to gray-box quadrotors without adding unnecessary complexity or additional control loops to the system. Two trimming methods are applied using Deep Reinforcement Learning and through the concept of Digital Twin in parallel. One of them is novel and based on reverse engineering, and another one is based on previous works, in which by comparing both methods, our method has faster convergence and efficiency.