GaussianWAM

GaussianWAM is a training-time 3D Gaussian enhancement framework for World-Action Models (WAMs). It distills geometry and semantics from a 3D Gaussian field into WAM representations while keeping the original inference path unchanged.

Project page: https://tuojingai.github.io/GaussianWAM-project-page/
arXiv: https://arxiv.org/abs/2608.24714

What it does

• Builds an offline 3D Gaussian teacher from synchronized multi-view observations
• Renders semantic, depth, and alpha targets onto WAM-aligned grids
• Distills cached Gaussian targets into current-observation visual representations
• Removes all Gaussian-related modules at inference

Why it helps

GaussianWAM unifies geometry, visual semantics, and rendering-derived coverage in one spatial carrier. The same training-time supervision can be applied to both FastWAM-style dual-expert MoT models and Cosmos-Policy-style unified Transformers.

Results

• LIBERO: 98.6% average success
• LIBERO-Plus: 87.8% overall
• Real-world FastWAM: 30.0% -> 40.0%

Citation

@article{gaussianwam2026,
  title   = {GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models},
  author  = {Zhang, Zijian and Jiang, Yuqing and Zhou, Weitao and Li, Minglei and Zhang, Jinhao and Mu, Yao and Li, Xiaofan and Zhao, Hao and Yu, Haibao},
  journal = {arXiv preprint arXiv:2608.24714},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.24714}
}