Gaussian Process Constraint Learning for Scalable Chance-Constrained Motion Planning From Demonstrations

We propose a method for learning constraints represented as Gaussian processes (GPs) from locally-optimal demonstrations. Our approach uses the Karush-Kuhn-Tucker (KKT) optimality conditions to determine where on demonstrations constraint is tight, and scaling of gradient at those states. then train GP representation which consistent with generalizes this information. further show that uncertainty can be used within kinodynamic RRT plan probabilistically-safe trajectories, we exploit structure planner exactly achieve specified safety probability. demonstrate our learn complex, nonlinear demonstrated 5D nonholonomic car, 12D quadrotor, 3-link planar arm, all while requiring minimal prior information constraint. results suggest learned accurate, outperforming previous methods require more a priori knowledge.