We introduce a novel offline RL algorithm that leverages shortcut models to scale both training and inference.

We show that diffusion policy can be made faster and more accurate without retraining and analyse why, we introduce a genetic algorithm to reduce to two denoising steps..