RESEARCH PAPERYear 2025

Latent Action Pretraining from Videos

Seonghyeon Ye; Joel Jang; Byeongguk Jeon; Sejune Joo; Jianwei Yang; Baolin Peng; Ajay Mandlekar; Reuben Tan; Yu-Wei Chao; Bill Yuchen Lin; Lars Liden; Kimin Lee; Jianfeng Gao; Luke Zettlemoyer; Dieter Fox; Minjoon Seo

Read paper PDF Code Project page

Classification

View four quadrants

Major category: Components of WAMs
Quadrant: Not applicable
Architecture: Not applicable
Prediction paradigm: Not applicable
Subcategories: Action representations & policies Latent action pretraining
Source review status: Verified from primary sources

Category review. LAPA learns an explicit quantized inverse-transition representation that labels actionless video, then uses these latent-action targets to pretrain a separate VLA. The manuscript explicitly identifies this latent action model as an action-representation component. Retention refers to that reusable LAM, not a claim of online world-model planning. Reading evidence

AT A GLANCE

Contribution

A contribution summary has not been added yet.

Abstract

An abstract has not been added yet.

Affiliations

Not listed in the collection.

BibTeX

@inproceedings{lapa,
  author       = {Seonghyeon Ye and
                  Joel Jang and
                  Byeongguk Jeon and
                  Se June Joo and
                  Jianwei Yang and
                  Baolin Peng and
                  Ajay Mandlekar and
                  Reuben Tan and
                  Yu{-}Wei Chao and
                  Bill Yuchen Lin and
                  Lars Liden and
                  Kimin Lee and
                  Jianfeng Gao and
                  Luke Zettlemoyer and
                  Dieter Fox and
                  Minjoon Seo},
  title        = {Latent Action Pretraining from Videos},
  booktitle    = {ICLR},
  year         = {2025},
}

Classification

Contribution

Abstract

Affiliations

BibTeX

More in this category

V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning

DINOv3

World Simulation with Video Foundation Models for Physical AI