Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control
NowWAM argues that denoising the current observation is enough to adapt generative DiTs for robot control.
The paper says future visual prediction is not essential under matched controlled settings, because past and future targets performed comparably. Training only on the clean endpoint hurt robustness, pointing to the denoising trajectory as the useful interface. On LIBERO-Plus, NowWAM reached 87.7% with FLUX2-Klein, beating the future-target baseline by 6.1 points while cutting visual tokens and step time. With a pure text-to-image Z-Image backbone, it reached 87.8%, suggesting the method is not dependent on video-generation or image-editing models. HF Daily Papers' note
The paper says future visual prediction is not essential under matched controlled settings, because past and future targets performed comparably. Training only on the clean endpoint hurt robustness, pointing to the denoising trajectory as the useful interface. On LIBERO-Plus, NowWAM reached 87.7% with FLUX2-Klein, beating the future-target baseline by 6.1 points while cutting visual tokens and step time. With a pure text-to-image Z-Image backbone, it reached 87.8%, suggesting the method is not dependent on video-generation or image-editing models. HF Daily Papers' note
score 5