This is the official repo for the ACL 2026 Main Conference paper NaviMaster: Learning a Unified Policy for GUI and Embodied Navigation Tasks.
[🏠Homepage] | [💻Code] | [📝Paper] | [🤗Models] | [🤗Data]
- [2026/04/07] NaviMaster is accepted to ACL 2026 Main Conference!
- [2026/03/18] We released the code and scripts.
- [2026/01/14] We released Dataset!
- [2025/11/08] We released Model!
- [2025/08/04] Our NaviMaster paper (NaviMaster: Learning a Unified Policy for GUI and Embodied Navigation Tasks) can be accessed in arXiv!
NaviMaster is the first unified agent that jointly handles Graphical User Interface (GUI) navigation and embodied navigation within a single framework. By reformulating both tasks into a visual-target trajectory format and training a unified policy via reinforcement learning (GRPO) with a distance-aware dense reward, NaviMaster achieves strong cross-domain generalization and outperforms state-of-the-art specialists on out-of-domain benchmarks.
NaviMaster/
├── assets/ # Figures and teaser images
├── data/ # Training data
├── verl/ # Core code
│ ├── models/ # Model wrappers (Qwen2.5VL, GRPO policy)
│ ├── single_controller/ # Controller logic for resource and execution flow
│ ├── trainer/ # GRPO trainer (adapted from EasyR1)
│ ├── utils/ # Utils
│ └── workers/ # Distributed worker processes
├── scripts/ # Entry points
│ ├── train.sh # Unified training with GRPO
│ ├── runtime_env.yaml
│ ├── model_merger.py # Merge Checkpoints
│ └── config.yaml # Training/Val configs (YAML)
├── tests/ # Evaluation
│ ├── test-embodied # Evaluation on Embodied benchmarks
│ └── test-gui # Evaluation on GUI benchmarks
├── requirements.txt # Python dependencies
├── setup.py # Package installation
└── README.md # This file
git clone https://github.com/Iron-Boyy/NaviMaster.git
cd NaviMaster
pip install -r requirements.txt
Download NaviMaster-Dataset-2w and place it under data/.
Train NaviMaster on mixed GUI + embodied data:
cd scripts/
bash train.sh
First prepare the evaluation dataset Android Control and place it under test-gui/data/androidcontrol. Then run the following code:
cd tests/test-gui
bash inference.sh
bash eval.sh
First prepare the evaluation dataset Where2Place, RoboSpatial, RoboReflt, RefSpatial and place it under test-embodied/data/. Then run the following code:
cd tests/test-embodied
bash run_test.sh
bash eval.sh
The red background represents that the data source is in the training set of the corresponding model, while the green background represents that the test dataset is OOD for the model. Bold highlights the best results in the OOD setting, and underlined are the second-best.
These results demonstrate that NaviMaster’s fine- grained visual–spatial alignment significantly en- hances performance in both object-level and free- space referring. Interpolate start reference image.
NaviMaster achieves the highest Success Rate and SPL, representing a substantial improvement over the base model.
If this repository or our dataset contributes to your research, please acknowledge it by citing our paper as follows.
@article{luo2025navimaster,
title={Navimaster: Learning a unified policy for gui and embodied navigation tasks},
author={Luo, Zhihao and Yan, Wentao and Gong, Jingyu and Wang, Min and Zhang, Zhizhong and Wang, Xuhong and Xie, Yuan and Tan, Xin},
journal={arXiv preprint arXiv:2508.02046},
year={2025}
}