Hugging Face's logo

ldx825
/
WAM-Adapter

Robotics
PyTorch
robot-manipulation
vision-language-action
world-models
video-diffusion
diffusion-transformer
lora
imitation-learning
robotwin
libero
Model card Files
xet
 Community
WAM-Adapter
File size: 3,369 Bytes
a46aaf1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f26de1a
 
 
 
a46aaf1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dda6ad2
a46aaf1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---

license: mit
library_name: pytorch
tags:
  - robotics
  - robot-manipulation
  - vision-language-action
  - world-models
  - video-diffusion
  - diffusion-transformer
  - lora
  - imitation-learning
  - robotwin
  - libero
---


# WAM-Adapter

**WAM-Adapter: Efficiently Adapting Video-DiT World Representations to Robot Actions**

Dongxu Liu<sup>1,2,*</sup>&ensp; Zhe Li<sup>1,*</sup>&ensp; Zundong Ke<sup>1</sup>&ensp; Zhipeng Zhang<sup>1,†</sup>

<sup>1</sup>AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University<br>
<sup>2</sup>Awaken Robotics<br>
(*Equal contribution, †Corresponding author)



## 📖 Overview



WAM-Adapter is a parameter- and training-efficient framework that adapts a **frozen Video-DiT

world model** (`Wan2.1-T2V-1.3B`) into a robot *World Action Model*:



- **Frozen Video-DiT + LoRA.** The 1.3B backbone stays frozen; rank-64 LoRA adapts it to robot data.

- **One joint current+future pass.** The clean current frame and noisy future latent slots are

  processed in a *single* Video-DiT forward — no future-video denoising loop, no extra action-side
  forward.
- **Temporal-group cross-layer readout.** Multi-layer representations are organized into
  **Current / Early-Future / Late-Future** groups with group-specific cross-layer fusion, and the
  whole action chunk is predicted by a small mixed-context action decoder.

Only **121M trainable parameters** (with a compact 1.3B backbone, no embodied pretraining):
**89.0%** RoboTwin 2.0 (50 tasks), **98.5%** LIBERO, **75.17%** LIBERO-Plus — while improving
training throughput by **9.66×** over Fast-WAM under the same setting.

## 📦 Checkpoints

| File | Benchmark | Training steps | Size |
|---|---|---|---|
| `wam_adapter_robotwin50_240k.pt` | RoboTwin 2.0 (50 tasks) | 240k | 3.08 GB |
| `wam_adapter_libero_spatial_70k.pt` | LIBERO-Spatial | 70k | 3.08 GB |
| `wam_adapter_libero_object_60k.pt` | LIBERO-Object | 60k | 3.08 GB |
| `wam_adapter_libero_goal_80k.pt` | LIBERO-Goal | 80k | 3.08 GB |
| `wam_adapter_libero_long_80k.pt` | LIBERO-Long | 80k | 3.08 GB |

Each checkpoint is a PyTorch file containing the LoRA-adapted backbone (`mot`), the temporal-group
readout (`action_decoder`) and its configuration.

## 🚀 Usage

```python

from huggingface_hub import hf_hub_download



ckpt_path = hf_hub_download(

    repo_id="ldx825/WAM-Adapter",

    filename="wam_adapter_robotwin50_240k.pt",

)

# then, inside the WAM-Adapter codebase:

# model = ...  # build WAMAdapter (see the GitHub repository)

# model.load_checkpoint(ckpt_path)

```

Please refer to the [GitHub repository](https://github.com/ldx825/WAM-Adapter) for the complete
training / evaluation / deployment instructions.

## 🔗 Links

- **GitHub**: https://github.com/ldx825/WAM-Adapter
- **Project page**: https://ldx825.github.io/WAM-Adapter-Homepage/
- **Backbone**: [Wan-AI/Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B)

## 📝 Citation

The arXiv release is coming soon; for now please cite the project as:

```bibtex

@article{wamadapter2026,

  title   = {WAM-Adapter: Efficiently Adapting Video-DiT World Representations to Robot Actions},

  author  = {Liu, Dongxu and Li, Zhe and Ke, Zundong and Zhang, Zhipeng},

  journal = {arXiv preprint},

  year    = {2026},

}

```