Control Barriers in Bayesian Learning of System Dynamics

This article focuses on learning a model of system dynamics online, while satisfying safety constraints. Our objective is to avoid offline identification or hand-specified models and allow safely autonomously estimate adapt its own during operation. Given streaming observations the state, we use Bayesian obtain distribution over dynamics. Specifically, propose new matrix variate Gaussian process (MVGP) regression approach with an efficient covariance factorization learn drift input gain terms nonlinear control-affine system. The MVGP then used optimize behavior ensure high probability, by specifying control Lyapunov function (CLF) barrier (CBF) chance We show that safe policy can be synthesized for systems arbitrary relative degree probabilistic CLF-CBF constraints solving second-order cone program. Finally, extend our design self-triggering formulation, adaptively determining time at which needs applied in order guarantee safety.