Optimal Control and Robot Learning on Agile Safety-Critical Systems

Abstract

We present a pipeline of optimal control methods to learn an optimal control policy and locally accurate dynamics models for agile and safety-critical robots using autonomous racing as an application example. We introduce Spline-Opt, a fast offline/online optimization and planning method that can produce a reasonably good initial optimal trajectory given very little dynamics data. We then introduce EL-MPC, a residual learning MPC that relies on prior data to estimate dynamics models and learn the optimal control policy bounded by safe set constraints. All together, the data-driven pipeline is a road map going from zero understanding of a robot's dynamics, to the mastery of its handling limit and optimal performance.