Synthesising Controllers for Quadrotors via Imitation Learning

In this work, we present a quadrotor control solution using neural networks with imitation learning. Instead of mapping raw sensory variables to control outputs, we leverage a feature engineering method to design neural network input attributes, according to expert controller inputs. This improves the capability of the learning-based control policies on tracking different trajectories. Moreover, the neural network outputs of the collective thrust and torque can be directly fed to the quadrotor control allocator, without requiring any further data transformation. During the training process, we also use the Data Aggregation (DAgger) algorithm for the improved control performance. The simulation results show that the quadrotors, using the trained neural network controller, can achieve expert-like behaviour (i.e. tracking square trajectories).