lawam_pretrain

LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

LaWAM (Latent World Action Model) is a robotics policy model that exposes predictive dynamics to robot policies through compact latent visual subgoals instead of reconstructed future video. It achieves state-of-the-art or competitive success rates across LIBERO, RoboTwin, and real-world manipulation tasks while maintaining low-latency inference.

• Paper: LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies
• Project Page: LaWAM Project Page
• Repository: GitHub - RLinf/LaWAM

Method Overview

LaWAM introduces a latent world-model interface for VLA policies. It predicts future observation features in a frozen visual feature space and injects them as latent visual subgoals for action generation.

<p align="center"> <img src="https://raw.githubusercontent.com/RLinf/LaWAM/main/assets/lawam_overview.png" alt="LaWAM method overview" width="95%"> </p>

Setup and Inference

Please refer to the official GitHub repository for instructions on environment setup, dataset preparation, SFT training, and running inference on the LIBERO and RoboTwin simulators.

Citation

If you find LaWAM useful in your research, please consider citing:

@misc{chen2026lawam,
  title = {LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies},
  author = {Chen, Jialei and Wang, Kai and Chen, Kang and Chen, Shuaihang and Gao, Feng and Tang, Wenhao and Li, Zhiyuan and Liu, Weilin and Yao, Zhuyu and Li, Boxun and Xu, Yuanbo and Yu, Chao},
  journal = {arXiv preprint arXiv:2606.15768},
  year = {2026},
  archiveprefix = {arXiv},
  primaryclass = {cs.RO},
}