Safety-critical control of nonlinear systems using deep neural network-based control barrier functions

Abstract

Safety-critical control of nonlinear systems based on control barrier functions is challenging, since uncertainties may invalidate the associated safety guarantees. To address this issue, an adaptive safety controller is developed by integrating a deep neural network with a tuneable input-to-state safe high-order control barrier function. In particular, the network is trained online to estimate uncertainty through an observer-driven hybrid learning mechanism, where stochastic gradient descent updates the hidden layers and recursive least squares updates the output-layer weights. The uncertainty estimate is incorporated for compensation, and forward invariance of an extended safe set is established via Nagumo's theorem under the proposed framework. Experimental results on a Franka Emika Panda robot demonstrate improved safety performance compared with existing baselines.