Deep Learning a Quadrotor Dynamic Model for Multi-Step Prediction

In this work, we develop and assess models for a real quadrotor in a Multi-Step prediction problem, that is, predicting the system state over many future steps using only the input. We propose a hybrid model with two configurations by combining deep recurrent neural networks with a quadrotor motion model. The proposed models take only motor speeds as input and predict the system state over a prediction length as long as 2 seconds. Network state initialization is performed and the effect of state initialization is shown in our results. Our experiments, evaluated on our collected real quadrotor dataset, show that the proposed models outperform the rigid body dynamics model, commonly used in quadrotor control, as well as a black-box model in both single-step and multi-step prediction scenarios.