Hybrid Dynamic Modelling using FeedForward and Temporal Convolutional Networks (FNN+TCN) and Robust Control Scheme for Aerial Manipulators
Abolfazl Eskandarpour, Mohammad Soltanshah, Kamal Kant Gupta, Mehran Mehrandezh · 2024
The dynamic complexity of unmanned aerial manipulators (UAMs), due to cross-coupled dynamics between the unmanned aerial vehicle (UAV) and the manipulator, challenges real-time optimal controller design. We present a novel hybrid framework for UAM modeling and optimal control for trajectory tracking using collected platform data. After analytically modeling the cross-coupled dynamics, we propose a feedforward neural network (FNN) architecture to identify the UAV’s nominal model and the cross-coupled matrices that represent the dynamic coupling between the UAV and the arm. Furthermore, we model uncertainty behaviors using a deep quantile regression model that provides explicit probability distributions at each sampling time. For this step, a deep temporal convolutional network (TCN) is used in which current and delayed states of the systems are incorporated into the model to obtain a more accurate uncertainty function, facilitating a less conservative tube-based robust controller design. Utilizing models from the FNN+TCN architecture, we propose a control scheme with linear constrained model predictive control (MPC) for translational dynamics and tube-based MPC for rotational dynamics, ensuring stability and robustness. We validate our framework’s performance through simulations and experiments.