jialei02/lawam_lam

dino_large_vae LaWAM Release

This directory is the minimal release package needed by LaWAM training.

This model is part of the work presented in the paper LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies.

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

Kept files:

• dino_large_vae.yaml: LAM model configuration consumed by lam_yaml_path.
• checkpoints/pytorch_model.pt: LAM release checkpoint consumed by lam_ckpt_path. It keeps only the state_dict needed by LaWAM training and inference.

Example LaWAM config:

framework:
  action_model:
    lam_ckpt_path: "latent_action_model/logs/dino_large_vae/lam_release/checkpoints/pytorch_model.pt"
    lam_yaml_path: "latent_action_model/logs/dino_large_vae/lam_release/dino_large_vae.yaml"

Citation

@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},
}