Recurrent Convex Difference Neural Networks for Safety-Critical Model Predictive Control

Abstract

Optimal control and planning with safety considerations constitute a fundamental challenge in model predictive control (MPC) applications, which has recently been addressed by integrating Control Barrier Functions (CBFs) to yield a safety-critical form of MPC, known as MPC-CBF. However, current neural network based approaches often face slow convergence speeds and limited prediction capabilities for MPC-CBFs. To address these limitations, we propose a Recurrent Convex Difference Neural Network (RCDiNN) framework, which can efficiently balance the prediction accuracy and convergence speed for model predictive control. It first incorporates RCDiNN to predict system dynamics and optimize control actions, and then employs a Lagrangian dual deep learning method for RCDiNN training, to encourage the satisfaction of the constraints given by MPC-CBF. We conduct an experimental evaluation on several benchmarks for obstacle avoidance, which demonstrates that our approach is more effective than the existing neural network-based MPC approaches.