Researchers have developed NowWAM, a novel co-training method for robot control that leverages pretrained generative Diffusion Transformers (DiTs). Unlike previous approaches that rely on future visual prediction, NowWAM directly couples the DiT's native generative objective to action prediction by denoising the current observation. This method demonstrates improved robustness and efficiency, achieving 87.7% accuracy on the LIBERO-Plus benchmark with a FLUX2-Klein backbone, a 6.1-point improvement over existing methods. Furthermore, NowWAM significantly reduces training time and visual tokens required, and shows strong performance even with a pure text-to-image backbone like Z-Image. AI
IMPACT This research could lead to more efficient and robust robot control systems by better adapting large generative models for real-world tasks.
RANK_REASON Academic paper detailing a new method for robot control using generative models. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
