SOTA methods for Imitation Learning are heavily data-inefficient or rely on unstable Adversarial training. We present a novel framework that introduces simple, stable, and data-efficient learning from few experts that scales well to complex environments.
