Event

EMMA Enhances VLA Policy Generalization with Generated Visual Data

1 reports · 1 sources

Zhehao Dong et al. propose the EMMA framework, which uses a generative data engine and training pipeline to augment training data for vision-language-action (VLA) policies, alleviating the high cost of collecting robot manipulation data across appearances. Its core component, DreamTransfer, is based on a diffusion Transformer and generates multi-view-consistent, geometrically plausible manipulation videos; it can perform text-driven editing of foreground, background, and lighting while preserving 3D structure. The paper reports that under zero-shot visual settings on real robot tasks, the framework achieves over 92% relative performance improvement compared with training on real data only, and adding the AdaMix training strategy yields an additional 17% improvement; AdaMix adaptively weights samples according to policy performance, focusing on harder samples.