jialei02/libero_merged_no_noops_20hz

LaWAM Dataset

This repository contains dataset artifacts from the paper LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies.

• Project Page: https://rlinf.github.io/LaWAM/
• Repository: https://github.com/RLinf/LaWAM

Dataset Description

LaWAM (Latent World Action Model) introduces a latent world-model interface for Vision-Language-Action (VLA) policies. It exposes predictive dynamics to robot policies through compact latent visual subgoals instead of reconstructed future video, enabling efficient and dynamics-aware robot control.

This dataset includes preprocessed trajectories converted to the LeRobot 3.0 format, which are used for training and evaluating LaWAM on robot manipulation tasks (such as LIBERO and RoboTwin).

Citation

If you find this dataset or the codebase useful, please cite the paper:

@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},
}