System hardware of UAV

The hardware adopts a two-axis variable-pitch UAV platform: the fuselage is made of 5mm carbon fiber plates to realize light weight and vibration reduction, the pitch control mechanism is based on helicopter tail rotor design, the steering gear is placed near the center of the fuselage and drives the blades through the push rod (reducing the moment of inertia), and all the connectors are made by 3D printing. The control system uses STM32 microcontroller and runs FreeRTOS real-time operating system: reinforcement learning controller implements lightweight strategy network (neural network with 96+64 nodes) at 100Hz frequency, and processes gyro filter data at 1000Hz. The experimental environment relies on the Vicon motion capture system to provide full state feedback (position/attitude/speed/angular velocity), and is equipped with a double parallel linear slide rail restraint device to limit the UAV motion in the vertical plane to verify the zero sample migration performance of the reinforcement learning control strategy.

We realized the efficient migration from simulated training to real deployment, and completed challenging flight actions on the physical platform. The practical hardware design framework based on reinforcement learning proposed in this work is shown in the following figure.

The top layer outputs the target angular velocity for strengthening the learning strategy network;

The middle layer is a high-frequency PD controller for fine attitude adjustment;

The bottom layer is an adaptive variable pitch thrust control module, which realizes fast and accurate bidirectional thrust output;

At the same time, a system structure with the consistency of reality and reality is constructed to bridge the system differences between training and deployment.