Computing Sound and Accurate Upper and Lower Bounds on Hamilton-Jacobi Reachability Value Functions

Abstract: Hamilton-Jacobi (HJ) reachability analysis is a fundamental tool for safety verification and control synthesis for nonlinear-control systems. Classical HJ reachability analysis methods discretize the continuous state space and solve the HJ partial differential equation over a grid, but these approaches do not account for discretization errors and can under-approximate backward reachable sets, which represent unsafe sets of states. We present a framework for computing sound upper and lower bounds on the HJ value functions via value iteration over grids. Additionally, we develop a refinement algorithm that splits cells that were not possible to classify as safe or unsafe given the computed bounds. This algorithm enables computing accurate over-approximations of backward reachable sets even when starting from coarse grids. Finally, we validate the effectiveness of our method in two case studies.