NowWAM: Future-Target-Free Co-training for Robot Control
September 22, 2026
NowWAM adapts pretrained Diffusion Transformers (DiTs) for robot control by co-training on denoising current observations and predicting actions. This approach avoids the need for future visual prediction, directly coupling the generative objective to action-facing representations.
HOW THIS AFFECTS YOU
●
builderYou may find more efficient ways to adapt large-scale video models for low-latency robotics.
●
researcherThis explores the fundamental utility of generative priors in downstream control tasks.