Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control

Beyond Future Prediction. Denoising as Generative Adaptation for Robot Control. By Zanyi Wang and six others, from UC San Diego, the University of Hong Kong, and two other institutions. Published on the twenty-third of September. Its code repository is public and real: thirteen source files.

Pretrained Diffusion Transformers, or DiTs, pick up rich pixel-level visual structure, and structure conditioned on language, from large-scale image and video generation training. A growing line of robot policies builds on that generative prior. How the prior should be transferred to control, though, remains unclear. Existing approaches commonly do it through future visual prediction. The authors step back to a more basic question. What does a pretrained generative DiT actually contribute to action learning, and how should that prior be adapted for control?

Their answer is NowWAM, a co-training formulation with no future target at all. Instead of predicting what comes next, it denoises the current observation, and it predicts robot actions from the same visual stream. That couples the model's native generative objective directly to the representation facing the actions, and it does so across the whole denoising trajectory.

Controlled comparisons probe what matters in that setup. Under matched settings, past and future visual targets perform comparably. Restricting training to the clean endpoint of denoising, by contrast, substantially reduces robustness. Read together, the two results suggest that a separate future target is not essential for generative adaptation, while the denoising trajectory remains an effective interface for control. The useful part, on this evidence, is the denoising itself rather than looking ahead.

On LIBERO-Plus, NowWAM with FLUX2-Klein reaches eighty-seven point seven percent. That is six point one points above the co-training baseline that uses a future target, and it comes while halving the training visual tokens, from seven hundred eighty-four down to three hundred ninety-two. Step time drops from two point eight five seconds to one point six three seconds, a one point eight times speedup.

The backbone does not have to come from video, either. With Z-Image, a pure text-to-image backbone, NowWAM reaches eighty-seven point eight percent, which the authors read as showing that strong control adaptation is not tied to video generation or image-editing backbones.

Existing approaches commonly transfer a generative prior to a robot by having it predict the future. NowWAM tests whether that is needed and finds the future target can be dropped. Denoising the present, and reading actions off the same stream, keeps what the prior offers, uses half the visual tokens, and trains faster per step.