TagaVLM: Topology-Aware Global Action Reasoning for Vision-Language Navigation

Official implementation of the ICRA 2026 paper "TagaVLM: Topology-Aware Global Action Reasoning for Vision-Language Navigation".

For details, please visit our project page.

Results on R2R (Val Unseen)

Method	Backbone	NE ↓	OSR ↑	SR ↑	SPL ↑
NavCoT	LLaMA2-7B	6.26	48.11	40.23	36.64
MapGPT	GPT-4V	5.62	57.9	47.7	38.1
TagaVLM-0.5B (Ours)	Qwen2-0.5B	5.57	55.09	45.72	41.91
TagaVLM-7B (Ours)	Qwen2-7B	4.97	60.2	51.09	47.18

Installation

Requires uv and Python 3.9-3.11.

git clone https://github.com/APEX-BJUT/Taga-VLM.git
cd Taga-VLM

# Inference only
uv sync

# Training (includes deepspeed, wandb, peft, etc.)
uv sync --extra train

This will create a .venv, install all dependencies, and build the patched transformers (required for STAR-Att) automatically.

Flash-Attention 2 (optional): Download the prebuilt .whl for your CUDA/Python version from Flash-Attention Releases (select the abiFALSE variant), then:

uv pip install flash_attn-*.whl

Matterport3D Simulator: Follow Matterport3DSimulator.

Data Preparation

Download model weights and data from HuggingFace:

# Model weights
huggingface-cli download tiredtony/TagaVLM-qwen2-0.5b --local-dir model_zoo/TagaVLM-qwen2-0.5b
huggingface-cli download tiredtony/TagaVLM-qwen2-7b   --local-dir model_zoo/TagaVLM-qwen2-7b

# Dataset
huggingface-cli download tiredtony/TagaVLM_infer_data --repo-type dataset --local-dir data

Expected directory structure:

Taga-VLM/
├── data/
│   ├── R2R/
│   │   ├── annotations/
│   │   └── connectivity/
│   ├── mp3d_data/
│   ├── view_images_bgr_from_mattersim.h5
│   ├── view_images_hm3d/
│   └── anno/
├── model_zoo/
│   ├── TagaVLM-qwen2-0.5b/
│   └── TagaVLM-qwen2-7b/

Training & Evaluation

Training

bash scripts/train/finetune_TagaVLM.sh

Note: For the 0.5B model, add "vocab_size": 151936 and "tie_word_embeddings": true to config.json after training.

Evaluation

cd map_nav_src && bash run_r2r.sh

To switch between models, edit model_zoo/TagaVLM-qwen2-* path in map_nav_src/r2r_llava/agent_base.py. The spatial pool stride is read from each model's config.json (mm_spatial_pool_stride: 3 for 0.5B, 2 for 7B).

Citation

@inproceedings{liu2026tagavlm,
  title     = {TagaVLM: Topology-Aware Global Action Reasoning for Vision-Language Navigation},
  author    = {Liu, Jiaxing and Zhang, Zexi and Li, Xiaoyan and Wang, Boyue and Hu, Yongli and Yin, Baocai},
  booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}

Acknowledgement

This project builds upon LLaVA-NeXT and VLN-DUET. We thank the authors for open-sourcing their code.