EMMA: Enhancing Robotic Control Robustness through a Generative Visual Model
EMMA introduces a framework combining DreamTransfer and AdaMix to improve vision-language-action (VLA) policies for robot manipulation. It generates multi-view consistent, geometrically plausible videos using DreamTransfer, allowing text editing of scenes. AdaMix is a training strategy that adaptively weights samples based on performance to emphasize challenging data, leading to over 92% relative performance gains in real-world zero-shot tasks.