Bi-Weekly Talk: Henrik Hose: Approximate Model Predictive Control with Guarantees
Wednesday, October 26, 2022, 10:30am
Location: RWTH Aachen University, Department of Computer Science - Ahornstr. 55, building E3, room 9u10
Speaker: Henrik Hose
Approximate Model Predictive Control with Guarantees
Fast feedback responses, stability, and constraint satisfaction are critical requirements for control in robotics to ensure safety. Model predictive control (MPC) achieves stability and constraint satisfaction, but is notoriously slow to evaluate. Approximation of such MPC controllers via (deep) neural networks (NNs) allows for fast online evaluation. However, the approximation introduces inaccuracies that can cause instabilities or constraint violations. In this work, we propose approximating the complete predicted input trajectory of the MPC with NNs. We extend existing statistical offline validation methods to provide guarantees for trajectories with high probability. Additionally, we propose a novel online validation method that provides hard guarantees for stability and constraint satisfaction in contrast to the usually used probabilistic guarantees. The method requires evaluation of the MPCs terminal controller and a single forward integration of the predicted inputs - both fast to compute on resource constrained robotic systems. The proposed control framework is illustrated on a standard nonlinear benchmark problem.