# mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs

**Authors:** Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, Elvis Nava  
**arXiv:** [2512.15692](https://arxiv.org/abs/2512.15692)  
**Date:** 2025-12-17  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2512.15692v1/

## Key points

- Current Vision-Language-Action Models (VLAs) in robotic manipulation rely on pretrained vision-language backbones from disjointed web data.
- This reliance places a heavy burden on the policy to understand physical dynamics and temporal dependencies solely from robot trajectories, requiring continuous large-scale expert data collection.
- A novel paradigm is proposed that integrates video into the pretraining process to jointly capture semantics and visual dynamics, isolating low-level control tasks for improved efficiency in robotic manipulation.
- The mimic-video: Video-Action Model (VAM) pairs a pretrained Internet-scale video model with an action decoder based on flow matching techniques conditioned on latent representations, functioning as an Inverse Dynamics Model (IDM).
- Extensive evaluations show that this new approach achieves state-of-the-art performance in both simulated and real-world tasks, enhancing sample efficiency by 10 times and accelerating convergence speed by 2 times compared to traditional VLA architectures.

## Summary

, , , , 

In the field of robotic manipulation, current Vision-Language-Action Models (VLAs) rely on pretrained vision-language backbones from disjointed web data. While this has improved semantic generalization, it places a heavy burden on the policy to understand physical dynamics and temporal dependencies solely from robot trajectories. This requires continuous large-scale expert data collection to compensate for the lack of inherent physical understanding. To address this limitation, a novel paradigm is proposed that integrates video into the pretraining process to jointly capture semantics and visual dynamics. This approach isolates low-level control tasks, leading to improved efficiency in robotic manipulation. The introduced model, known as mimic-video: Video-Action Model (VAM), pairs a pretrained Internet-scale video model with an action decoder based on flow matching techniques conditioned on latent representations. This decoder functions as an Inverse Dynamics Model (IDM), generating low-level robot actions from the latent representation of video-space action plans. Extensive evaluations demonstrate that this new approach achieves state-of-the-art performance in both simulated and real-world tasks, enhancing sample efficiency by 10 times and accelerating convergence speed by 2 times compared to traditional VLA architectures. The collaborative effort of authors Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava has paved the way for generalizable robot control beyond existing VLAs through the innovative mimic-video framework.