Born to generate. Adapted to act.

Beyond Future Prediction

Denoising as Generative Adaptation for Robot Control

The future is the target.
The representation is the point.

Read the paper ↗ BibTeX Code ↗

Zanyi Wang1 Yuheng Lei2 Dengyang Jiang3 Ping Luo2 Mengdi Wang4 Zhixuan Liang4,2 ✉ Shilong Liu4 ✉

1UC San Diego2HKU3HKUST4Princeton University

Paper overview: future prediction in the action loop, future-target co-training, and NowWAM current-stream denoising, with benchmark and training-efficiency results.

The idea in three steps

Learn to see through generation.
Adapt that representation to act.

01 / Previous work

Training · future-target co-training

Current demonstration frame 80: the robot holds the pear above the placemat.
Current
Future demonstration frame 96: the pear is above the bowl.
Future demonstration frame 112: the pear has been placed in the bowl.
Future demonstration frame 124: the hand moves away. Highlighted as a schematic single-frame target.
Future video

Prediction as adaptation.

A future target gives a pretrained generative model a familiar way to adapt its representations for control.

02 / Change the target

Change the time, not the task

Current
Past t − 16Future t + 32

83.80≈83.70

Past ≈ future.

Past and future targets perform comparably. The evidence points beyond forward prediction alone.

03 / NowWAM

Same current · different noise levels

One real current observation with changing illustrative training noise levels.

Sample → Train

Denoising + Action learning

Adapt the present.

NowWAM couples denoising and action learning on the current stream. No separate target needed.

Result

More robust.
Less overhead.

OOD control / LIBERO-Plus

87.7%

NowWAM · FLUX.2-Klein

ImageWAM 81.6%+6.1 pts

Training speed

≈1.8×

Faster steps. Half the visual tokens.

2.85 → 1.63 s / step

Text-to-image backbone

87.8%

Z-Image · LIBERO-Plus

Beyond video and editing backbones.
Robot shift

48.3 → 70.5%

Camera shift

77.7 → 88.6%

Standard LIBERO

98.4%

RoboCasa · 100 demos / task

64.9%

Citation

Build on
NowWAM.

@article{wang2026nowwam,
  title   = {Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control},
  author  = {Wang, Zanyi and Lei, Yuheng and Jiang, Dengyang and Luo, Ping and Wang, Mengdi and Liang, Zhixuan and Liu, Shilong},
  journal = {arXiv preprint arXiv:2609.28339},
  year    = {2026}
}