Multiple architectures.
One workflow.
5 architectures (Unified / Hidden / MoT / MoT-Joint / MoT-IDM) × 2 backbones (Wan2.2-5B / Cosmos2.5-2B).
WORLD ACTION MODELSOPEN RESEARCH
A unified research codebase
EasyWAM: A Unified and Efficient Framework for Training and Evaluating World Action Models
EasyWAM connects model implementation, data processing, distributed training and evaluation in one coherent workflow, so ideas travel from notebook to robot faster.
01 / overview
One framework.
Every layer in sync.
From sparse video decoding to resumable evaluation, EasyWAM makes the entire WAM loop inspectable and repeatable.
Multiple architectures.
2.37× Efficiency
Acceleration.
20,000-step LIBERO training drops from 14 hr to 6 hr compared to FastWAM.
Supports Full &
LoRA Fine-Tuning.
Full-parameter and LoRA fine-tuning, with LoRA training runnable on a single RTX 4090.
Benchmarks
Included.
Built-in evaluation infrastructure for LIBERO, LIBERO-Plus, RoboTwin and more.
02 / results
Strong results,
controlled experiment.
Wan2.2-TI2V-5B backbone · all scores are success rates (%), higher is better.
| Model | Spatial | Object | Goal | Long | Avg. |
|---|---|---|---|---|---|
| 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 |
| Model | Spatial | Object | Goal | Long | Avg. |
|---|---|---|---|---|---|
| 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 |
| 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 |
LIBERO-Plus perturbs background, camera, language, layout, lighting, noise and robot embodiment.
03 / blog
Living blog from
the WAM frontier.
Long-form analysis from the experiments behind EasyWAM.
04 / contributing
Let's build
something big.
EasyWAM grows through shared recipes, careful benchmarks and practical fixes. Whether you are adding a model, improving a dataloader or writing your first issue, there is a place for your work here.
Open an issue ↗01
Pick a path
Models · Data · Training · Evaluation
02
Build in public
Fork, branch, test and document the change.
03
Share the result
Open a pull request. We review with care.
$ git clone github.com/OpenMOSS/EasyWAMContact us and build EasyWAM together — start a discussion ↗