SuSIE generates subgoals using an image-editing diffusion model and executes those subgoals using a languages-agnostic low-level policy.
Since subgoal generateing process doesn’t require robot actions, they augment the BridgeData (only robot demostrations) with a human manipulation dataset (Something-Something) to train the diffusion model.