VLA-JEPA is a vision-language-action pretraining framework that learns action-relevant dynamics by predicting future latent states, followed by action-head fine-tuning. Paper authors are affiliated with USTC, Zhongguancun Academy, SJTU, Tsinghua, EIT Ningbo, UCAS and Nankai University.

Overview

Model type
Action policy
Listed versions
VLA-JEPA
Family introduction
Related robots
1

Model history

February 2026

VLA-JEPA introduces latent world-model training for robot control

New modelOfficial announcement

The first paper, published on February 10, introduces JEPA-style latent future-state supervision followed by action-head adaptation for robot control. Future frames are confined to target encoding. This adds a model and training option that combines human videos with action-labeled robot data. The first code or weight opening date is not established.

This announcement

The first paper, published on February 10, introduces JEPA-style latent future-state supervision followed by action-head adaptation for robot control. Future frames are confined to target encoding. This adds a model and training option that combines human videos with action-labeled robot data. The first code or weight opening date is not established.

Important because the newly disclosed training framework adds a practical model-selection and adaptation option using human-video supervision. This is not a claim of universal benchmark superiority; reported human-video effects are mixed.

View all important updates →

Papers & reports (1)

VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model

Paper available

arXiv: 2602.10098DOI: 10.48550/arxiv.2602.10098
Authors & affiliations (9)

Author order: Original order

  1. Jingwen SunEqual contribution

    University of Science and Technology of China · Zhongguancun Academy, Beijing, China

  2. Wenyao ZhangEqual contribution

    Shanghai Jiao Tong University · Eastern Institute of Technology, Ningbo

  3. Zekun Qi

    Tsinghua University

  4. Shaojie Ren

    Zhongguancun Academy, Beijing, China · University of Chinese Academy of Sciences

  5. Zezhi Liu

    Zhongguancun Academy, Beijing, China · Nankai University

  6. Hanxin Zhu

    University of Science and Technology of China

  7. Guangzhong Sun

    University of Science and Technology of China

  8. Xin JinCorresponding author

    Zhongguancun Academy, Beijing, China · Eastern Institute of Technology, Ningbo

  9. Zhibo ChenCorresponding author

    University of Science and Technology of China · Zhongguancun Academy, Beijing, China

First published
2026-02-10
Latest paper revision
v2 · 2026-02-14
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

Official resources

Submit a source or correction