Abstract: We propose a novel framework for data-efficient black-box robot learning under constraints. Our approach integrates probabilistic inference with Lagrangian optimization. With the guide of a ...