WNM-3D

WNM-3D is a generative world navigation model for continuous vision-language navigation. It converts monocular egocentric RGB history into persistent geometry-aware scene tokens and jointly generates future views and navigation actions for closed-loop control.

License notice: The released checkpoints embed VGGT-Ω/DINOv3-derived parameters. Use and redistribution are subject to the FAIR Noncommercial Research License, the DINOv3 License, and other applicable third-party terms.

• Project page
• Paper
• Source code and documentation
• GN0 / GN-Bench
• GN-Matrix dataset

Released Checkpoints

Directory	Training stage	Recommended use
wnm_3d_stage1_release	Offline A* SFT	Stage-I analysis and initialization
wnm_3d_stage2_release	Closed-loop DAgger-SFT	Stage-II analysis and initialization
wnm_3d_stage3_release	Counterfactual DanceGRPO	Evaluation and inference

The Stage-III checkpoint is the primary released policy. Each directory is a self-contained inference checkpoint; standalone Wan, UMT5-XXL, and VGGT-Ω initialization weights are not required for evaluation.

Download

hf download TeleEmbodied/WNM-3D \
  --include "wnm_3d_stage3_release/**" \
  --local-dir checkpoints

Inference

Install WNM-3D by following the installation guide, then launch a single policy replica:

bash scripts/inference/wnm_3d_server.sh \
  --model-path checkpoints/wnm_3d_stage3_release \
  --cuda-devices 0 \
  --num-replicas 1 \
  --base-port 8000

Run the GN-Bench client from a sibling GN0 checkout:

cd ../GN0
bash scripts/evaluation/eval_remote.sh \
  --exp-config configs/gn_bench/interiorgs/test_unseen.yaml \
  --enable-stall-recovery \
  --num-gpus 1 \
  --result-dir tmp/eval/wnm_3d_test_unseen

The reference single-replica deployment was validated on an NVIDIA H100 80 GB GPU and used approximately 27 GiB of GPU memory after loading. A released checkpoint occupies approximately 26 GiB. These figures are observations from the reference configuration, not strict minimum requirements.

Intended Use and Limitations

WNM-3D is intended for research on embodied navigation, world models, and closed-loop vision-language navigation. The released policy was developed and evaluated with the InteriorGS scenes and GN-Matrix task annotations used by GN-Bench. Performance may not transfer to new simulators, sensors, scene distributions, languages, or physical robots without additional validation.

Generated actions can fail or behave unexpectedly. Do not use the model as a safety-critical controller or deploy it around people, property, or physical systems without appropriate safeguards and human oversight.

License

The released checkpoints contain original WNM-3D parameters together with VGGT-Ω/DINOv3-derived parameters. They are not licensed solely under Apache-2.0. Use and redistribution are subject to all applicable terms, including the FAIR Noncommercial Research License and DINOv3 License.

The model repository provides the complete Apache-2.0, FAIR Noncommercial Research, and DINOv3 license texts, together with its Third-Party Notices. The source repository documents the complete component-level license boundaries. Datasets and separately downloaded initialization weights remain subject to their providers' terms.

Citation

@article{huang2026wnm_3d,
  title={WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN},
  author={Huang, Yuehao and Wu, Yunzi and Zhang, Xiaotao and Li, Xinhai and Dong, Jiankun and Lv, Jiajun and Zhang, Chi and Bai, Chenjia and Liu, Yong and Li, Xuelong},
  journal={arXiv preprint arXiv:2608.07267},
  year={2026}
}