DiffWAM-S
Small data. Freeze the Diff Block.
Overview
Jointly predict future states and consecutive visual differences.
World, Diff, and Action Experts learn under structured attention.
Deploy only the Observation Encoder and Action Expert.
Method
World and Diff Experts share early processing.
Predict future states.
Predict consecutive latent differences.
Current observation and each expert's own tokens.
No attention between different experts.
Keep the observation and action pathways only.
Training & inference
Small data. Freeze the Diff Block.
Large data. Pretrain, then train jointly.
Observation Encoder + Action Expert.
Results
+1.75 pts vs. FastWAM
+1.55 pts vs. FastWAM
+15.0 pts vs. FastWAM
98.8% average · +1.75 pts vs. FastWAM
| Method | Spatial | Object | Goal | Long | Avg. |
|---|---|---|---|---|---|
| Motus | 96.80 | 99.80 | 96.60 | 97.60 | 97.70 |
| Lingbot-VA | 98.50 | 99.60 | 97.20 | 98.50 | 98.45 |
| FastWAM | 98.20 | 99.80 | 97.80 | 92.40 | 97.05 |
| ImageWAM | 97.20 | 99.20 | 98.80 | 98.40 | 98.40 |
| WLA | 99.00 | 100.00 | 97.80 | 97.60 | 98.60 |
| DiffWAM-S (Ours) | 98.00 | 99.80 | 99.00 | 98.40 | 98.80 |
92.65% average · +1.55 pts vs. FastWAM
| Method | Clean | Random | Avg. |
|---|---|---|---|
| ABot-M0 | 86.60 | 85.08 | 85.84 |
| Motus | 88.66 | 87.02 | 87.84 |
| Lingbot-VA | 92.90 | 91.50 | 92.20 |
| FastWAM* | 91.88 | 90.32 | 91.10 |
| WLA | 92.94 | 90.02 | 91.48 |
| DiffWAM-L (Ours) | 93.28 | 92.02 | 92.65 |
| Method | Embodied pretraining | Clean | Random | Avg. |
|---|---|---|---|---|
| π0 | Yes | 46.4 | 16.3 | 31.35 |
| X-VLA | Yes | 68.0 | 20.9 | 44.45 |
| π0.5 | Yes | 70.7 | 46.0 | 58.35 |
| AHA-WAM | No | 64.3 | 3.2 | 33.75 |
| FastWAM† | No | 70.2 | 1.50 | 35.85 |
| DiffWAM (Ours) | No | 69.7 | 5.27 | 37.47 |
Citation
@article{zhang2026diffwam,
title = {DiffWAM: Learning World Action Model from Consecutive Visual Difference},
author = {Xinheng Zhang and Haoyu Xiong and Chao Jing and Xiaohan Zhang and Linshan Li and Hongbang Fan and Xingyu Chen},
journal = {arXiv preprint arXiv:2611.19904},
year = {2026},
eprint = {2611.19904},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.19904}
}