LaWAM LIBERO SFT Policy for LeRobot
This repository contains the released LaWAM LIBERO SFT policy converted to the
native LeRobot checkpoint format. Model parameters are stored in
model.safetensors; loading this repository does not require the original
PyTorch .pt checkpoint or local source-checkpoint paths.
- Paper: LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies
- Project page: rlinf.github.io/LaWAM
- Source code: RLinf/LaWAM
- Original checkpoint: jialei02/lawam_libero_sft_release
- LeRobot integration: huggingface/lerobot#3999
Loading
LaWAM support requires the LeRobot integration from PR #3999 or a later LeRobot release containing that integration.
from lerobot.policies.lawam import LaWAMPolicy
policy = LaWAMPolicy.from_pretrained("jialei02/lawam-libero-sft-lerobot")
The policy uses Qwen/Qwen3-VL-2B-Instruct as its portable base VLM identifier.
The Qwen model is downloaded automatically unless base_vlm_path is overridden
with a local copy.
The repository includes the policy config and complete LIBERO preprocessing and action postprocessing pipelines, including the released normalization tensors.
LIBERO Results
The released checkpoint was evaluated on all four LIBERO suites with 10 tasks per suite and 50 episodes per task. MuJoCo 3.3.2 is required to reproduce the released results exactly.
| Suite | Successes | Episodes | Success rate |
|---|---|---|---|
libero_spatial | 492 | 500 | 98.4% |
libero_object | 498 | 500 | 99.6% |
libero_goal | 490 | 500 | 98.0% |
libero_10 | 490 | 500 | 98.0% |
| Overall | 1970 | 2000 | 98.5% |
LIBERO results are sensitive to the MuJoCo version. Newer MuJoCo releases can reduce measured success rates for LaWAM and other policies; this is an evaluation environment compatibility issue rather than a checkpoint-format issue.
Conversion Verification
The source checkpoint, source config, dataset statistics, LAM architecture
config, and Qwen config/model were checked against their published SHA-256
values before conversion. A strict local from_pretrained() reload restored all
1,912 policy state keys while torch.load was disabled. A one-episode
libero_spatial inference smoke test completed successfully through the standard
lerobot-eval --policy.path path.
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},
}