EVA

This repository hosts the EVA checkpoint released with:

EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards

Project page: https://eva-project-page.github.io/
arxiv: https://arxiv.org/abs/2603.17808

Model Summary

This checkpoint is an EVA model for robotic video planning.

It is built on top of a Wan2.1 I2V 14B backbone and further adapted on Robotwin through:

• supervised fine-tuning (SFT)
• reinforcement-learning-based post-training (RL)

The released checkpoint corresponds to the merged model after both stages of post-training.

Intended Use

This model is intended for research use in robot video prediction and visual planning.

Given an input image and a language instruction, the model generates future video rollouts that are better aligned with executable robot behavior.