Fysics-AI/UniWAM

Paper · Project · GitHub · MAP-Bench

UniWAM unifies navigation and manipulation through a shared world-action backbone and Manipulation Anchor Pose (MAP) supervision. This repository provides the checkpoints and shared components for navigation on MAP-Bench:

• UniWAM-Base and UniWAM-Scaled, with inference configurations and normalization files.
• π₀.₅ navigation baseline checkpoint.
• Shared Wan2.2 VAE, UMT5-XXL text encoder, and tokenizer.

The released UniWAM checkpoints use a single navigation expert and one forward RGB view, since MAP-Bench evaluates navigation only. This navigation setting uses five sampled training frames and synchronous closed-loop execution; asynchronous committed-prefix execution is part of the full mobile manipulation path. The full mobile manipulation architecture and training/inference code are available on GitHub.

Download

Run from your workspace directory to download the complete model package (approximately 50 GB):

python -m pip install -U huggingface_hub
hf download Fysics-AI/UniWAM --local-dir ./UniWAM-weights

This includes both UniWAM checkpoints, the π₀.₅ comparison checkpoint, and the shared components. For selective downloads, use the GitHub download helper.

Usage

See GitHub for environment setup, inference and MAP-Bench evaluation, and training.

UniWAM weights use Apache-2.0. Bundled third-party models retain their upstream licenses.

Citation

@article{xue2026uniwam,
  title   = {UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision},
  author  = {Wei Xue and Keliang Liu and Mingzhang Cui and Jinhua Xie and Jinjie Wei and Jianan Hou and Jingcheng Lu and Lintao Wang and Kaixiang Qiu and Yizhou Liu and Xinghai Ye and Jinghang Han and Mingcheng Li and Jie Gu and Shunli Wang and Lihua Zhang and Dingkang Yang},
  journal = {arXiv preprint arXiv:2609.39388},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.39388}
}