Bridging the Sim-to-real Gap: A Control Framework for Imitation Learning of Model Predictive Control

원문 초록

To address the computational challenges of Model Predictive Control (MPC), recent research has studied on using Deep Neural Networks (DNNs) trained through imitation learning to approximate the MPC. However, this introduces a common issue in learning-based control: the simulation-to-reality (sim-to-real) gap. Therefore, Domain Randomization (DR) has been widely used to mitigate this gap by introducing perturbations in the source domain. However, this led to low data collection efficiency and an overly conservative control strategy. This study proposes a new control framework that deals with this issue from a control perspective inspired by Robust Tube MPC. The framework ensures the DNN operates in the same environment as the source domain, handling the sim-to-real gap with great data collection efficiency. Moreover, a parameter governor is introduced to address the DNN's inability to adapt to model parameter variations, enabling the system to satisfy MPC constraints more robustly under changing conditions. The proposed framework was validated through a cart-pole system case study compared by DR baselines, demonstrating that a single MPC-demonstrated trajectory in the source domain was sufficient for controlling the cart-pole in the target domain. Furthermore, the system effectively handled model parameter variations, allowing for a less conservative control.