Modeling Dynamic Systems for Multi-Step Prediction with Recurrent Neural Networks

This thesis investigates the applicability of Recurrent Neural Networks (RNNs) and Deep Learning methods for multi-step prediction of robotic systems. The unmodeled dynamics and simplifying assumptions in classic modeling methods result in models that yield rapidly diverging predictions when the model is used in an iterative fashion, i.e., for multi-step prediction. However, the effect of the unmodeled dynamics can be captured by collecting datasets of the system. Deep Learning provides a strong set of tools to extract patterns from data, however, large datasets are commonly required for the methods to work well. Collecting a large amount of data from a robotic system can be a cumbersome and expensive approach. In this work, Deep Learning methods, particularly RNNs, are studied and employed for the purpose of learning models of two aerial vehicles from experimental data. The feasibility of employing RNNs is first studied to learn a model of a quadrotor based on a simulated dataset, which yields a Multi Layer Fully Connected (MLFC) architecture. Models can be learned for multi-step prediction, recovering excellent predictions over 500 timesteps in the presence of simulated disturbances to the robot and noise on the measurements. To learn models from experimental data, the RNN state initialization problem is defined and formulated. It is shown that the RNN state initialization problem can be addressed by creating and training an initialization network jointly with the multi-step prediction network, and the combination can be used in a black-box modeling approach such that the model produces predictions which are immediately accurate. The RNN based black-box methods are trained on an experimental dataset gathered from a quadrotor and a publicly available helicopter dataset. The quadrotor dataset, which encompasses approximately 4