EasyWAM-MoT 采用独立的 Video DiT 与 Action DiT 专家,其 token 通过混合自注意力机制进行交互。该动作预测 checkpoint 以 Wan2.2-TI2V-5B 作为视频主干,并在 LIBERO 上进行了全量微调。
对应的 checkpoint 通过 EasyWAM 代码库训练获得。
在 EasyWAM LIBERO 评测协议下的成功率(%):
| Model | Spatial | Object | Goal | Long | Avg. |
|---|---|---|---|---|---|
| Full-Parameter | |||||
| EasyWAM-Unified | 99.0 | 99.4 | 99.2 | 98.2 | 99.0 |
| 🔥 EasyWAM-MoT | 97.8 | 98.4 | 97.6 | 95.6 | 97.4 |
| EasyWAM-Hidden | 99.4 | 100.0 | 97.0 | 97.8 | 98.6 |
| LoRA (Rank 128) | |||||
| EasyWAM-Unified | 84.0 | 97.8 | 92.0 | 81.2 | 88.8 |
| EasyWAM-MoT | 96.8 | 98.8 | 94.4 | 90.4 | 95.1 |
| EasyWAM-Hidden | 96.8 | 99.4 | 92.6 | 86.8 | 93.9 |
在 LIBERO-Plus 评测协议下的成功率(%):
| Model | Background | Camera | Language | Layout | Light | Noise | Robot | Avg. |
|---|---|---|---|---|---|---|---|---|
| EasyWAM-Unified | 55.8 | 33.7 | 93.7 | 80.6 | 92.2 | 50.2 | 71.4 | 67.5 |
| 🔥 EasyWAM-MoT | 52.8 | 20.6 | 80.4 | 65.2 | 85.1 | 51.5 | 49.7 | 56.8 |
| EasyWAM-Hidden | 56.8 | 49.2 | 95.3 | 81.0 | 90.4 | 58.2 | 77.4 | 72.4 |
hf download OpenMOSS-Team/EasyWAM-MoT-Wan22 \
easywam_mot_wan22.pt \
--local-dir ./checkpoints按照 EasyWAM LIBERO guide 中的说明,准备好 Wan2.2、LIBERO 以及匹配的 dataset_stats.json,然后运行:
python experiments/libero/run_libero_manager.py \
task=libero_easywam_mot_wan22 \
ckpt=./checkpoints/easywam_mot_wan22.pt \
EVALUATION.dataset_stats_path=<path-to-matching-dataset_stats.json>.pt)EasyWAM 代码以 MIT License 发布。使用该 checkpoint 亦须遵守其基础模型及训练数据的相关条款。如果 EasyWAM 对您的研究有所帮助,请引用:
@misc{easywam2026,
title = {EasyWAM: A Unified and Efficient Framework for Training and Evaluating World Action Models},
author = {EasyWAM-Team},
year = {2026},
url = {https://github.com/OpenMOSS/EasyWAM}
}