VAMPO：通过策略优化提升视频动作模型中的视觉动态

Video action models serve as an appealing foundation for Vision-Language-Action systems due to their ability to learn visual dynamics from large-scale video data and transfer this knowledge to downstream robot control. However, current diffusion-based video predictors are trained with likelihood-surrogate objectives that promote globally plausible predictions without explicitly optimizing the precision-critical visual dynamics necessary for manipulation. This objective mismatch often results in subtle errors in object pose, spatial relations, and contact timing, which can be amplified by downstream policies. VAMPO proposes a post-training framework that directly enhances visual dynamics in video action models through policy optimization. The core idea is to...