AI

Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 17:18:25 UTC.

Briefings

A new method called NowWAM uses a future-target-free co-training approach to adapt pretrained generative Diffusion Transformers for robot control. Experiments on LIBERO-Plus show it achieves 87.7% performance with FLUX2-Klein, outperforming future-target baselines by 6.1 points while reducing training visual tokens from 784 to 392 and cutting step time from 2.85s to 1.63s. The study concludes that a separate future target is not essential for generative adaptation, and the denoising trajectory serves as an effective interface for control.

Key points

• NowWAM is a future-target-free co-training formulation that denoises the current observation to predict robot actions directly from the same visual stream.
• On LIBERO-Plus, NowWAM reaches 87.7% performance with FLUX2-Klein, improving over future-target baselines by 6.1 points while halving training visual tokens from 784 to 392.

Context and limitations

Results are limited to LIBERO-Plus and FLUX2-Klein backbones; performance on pure text-to-image Z-Image backbone is reported at 87.8% but requires further validation.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/10/2026, 17:11:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeif73pmj5hkgbkcjpzcoym2knevn2tliwag5q3taqrpl5nfdln7ycm