Vision-language-action (VLA) models need diverse training data to generalize, but collecting large-scale robot manipulation data across varied object appearances is expensive and labor-intensive. The EMMA framework addresses this by pairing a generative data engine with a training pipeline that mixes real and synthetic data.
The generative component, DreamTransfer, is a diffusion Transformer architecture that produces multi-view consistent and geometrically plausible embodied manipulation videos. It supports text-driven editing of foreground, background, and lighting while preserving 3D structure and geometric fidelity. According to the paper, DreamTransfer outperforms previous video generation models on multi-view consistency, geometric fidelity, and text-conditioning precision.
For training, EMMA uses a hybrid set of real and generated data and introduces AdaMix, a strategy that adaptively weights samples according to policy performance to emphasize challenging ones. In real-world robotic tasks with zero-shot visual settings, the framework delivered a relative performance increase of over 92% compared to training with real data alone, and an additional 17% improvement when AdaMix was applied. The paper is a single source, so no conflicting findings are reported.