Safety-Critical Control Synthesis for Systems with Unknown Nonlinear Dynamics: An Adaptive Recurrent Neural Network Approach

This paper presents a safe control design for systems with unknown nonlinear dynamics, using a proposed adaptive recurrent neural network. For safety-critical control, several noteworthy results and methods have been presented that not only stabilize affine control systems to desired states but also ensure satisfaction of state and control constraints. One of the main challenges for safety-critical control design is the accurate knowledge of system dynamics, which is challenging to obtain for many practical systems subjected to limited computational resources and historical data. To address the fore-mentioned problem, we first present an adaptive recurrent neural network (AdaRNN) which leverages the concept of continual learning to learn and adapt to unknown nonlinear system dynamics with limited resources, followed by the utilization of AdaRNN outputs for the design of a control barrier function based quadratic program, to ensure safe behavior of the system. Finally, we illustrate the effectiveness of the presented approach through both field and simulation data.