[ICRA 2026] LaViRA: Language-Vision-Robot Actions Translation for Zero-Shot Vision-and-Language Navigation in Continuous Environments

TODO List

• Conda Environment Setup
• Evaluation Code
  • R2R-CE
  • RxR-CE
• Docker Environment Setup
• Real-world Deployment Code（Aloha & Go1）

Environment Setup

1. Create Conda Environment

The codebase has been tested with Python 3.9. Python 3.8 should also be compatible.

conda create -n lavira python=3.9
conda activate lavira

2. Install Habitat-Sim & Habitat-Lab

Habitat-Sim is compiled locally in this setup. Alternatively, a precompiled conda version is available.

Habitat-Sim

git clone https://github.com/facebookresearch/habitat-sim.git
cd habitat-sim
git checkout tags/v0.1.7
pip install -r requirements.txt
CMAKE_ARGS="-DCMAKE_POLICY_VERSION_MINIMUM=3.5" python setup.py install --headless

Habitat-Lab

Note: The tensorflow==1.13.1 dependency must be removed from habitat_baselines/rl/requirements.txt prior to installation to avoid dependency conflicts.

git clone https://github.com/facebookresearch/habitat-lab.git
cd habitat-lab
git checkout tags/v0.1.7
cd habitat_baselines/rl
vi requirements.txt  # Remove the line: tensorflow==1.13.1
cd ../../            # Return to habitat-lab root directory

pip install torch==2.1.0+cu121 torchvision==0.16.0 torchaudio==2.1.0+cu121 \
    -f https://download.pytorch.org/whl/torch_stable.html

pip install -r requirements.txt
# Installs both habitat and habitat_baselines.
# If installation fails due to network issues, please retry.
python setup.py develop --all

3. GroundingDINO

GroundingDINO is used to construct semantic maps for robot action planning.

git clone https://github.com/IDEA-Research/GroundingDINO.git
cd GroundingDINO
git checkout -q 57535c5a79791cb76e36fdb64975271354f10251
pip install -q -e . --no-build-isolation
pip install 'git+https://github.com/facebookresearch/segment-anything.git'
# Note: Upgrading setuptools and wheel may be required.

Phrase-to-Class Mapping Optimization

Following CA-Nav, we recommend replacing the default phrases2classes function in Grounded-SAM with a minimum edit distance-based implementation for more robust and stable prediction outputs.

pip install nltk

Locate the phrases2classes method at Line 235 of <YOUR_PATH>/GroundingDINO/groundingdino/util/inference.py. Comment out the original implementation and replace it with the following:

# @staticmethod
# def phrases2classes(phrases: List[str], classes: List[str]) -> np.ndarray:
#     class_ids = []
#     for phrase in phrases:
#         try:
#             class_ids.append(classes.index(phrase))
#         except ValueError:
#             class_ids.append(None)
#     return np.array(class_ids)

from nltk.metrics import edit_distance

@staticmethod
def phrases2classes(phrases: List[str], classes: List[str]) -> np.ndarray:
    class_ids = []
    for phrase in phrases:
        if phrase in classes:
            class_ids.append(classes.index(phrase))
        else:
            distances = np.array([edit_distance(phrase, c) for c in classes])
            class_ids.append(np.argmin(distances))
    return np.array(class_ids)

4. Other dependencies for LaViRA

Then install the remaining dependencies:

pip install setuptools==58.5.3 meson-python ninja
pip install -r requirements.txt --use-pep517 --no-build-isolation

---

Dataset Preparation

Download the Matterport3D scene dataset and the VLN-CE episode dataset, and place them under the data/ directory as described below.

Matterport3D Scenes

After obtaining the official download script (download_mp.py) by submitting the Terms of Use agreement at the Matterport3D project page, run:

python download_mp.py --task habitat -o data/scene_datasets/mp3d/

VLN-CE Episodes

Download R2R_VLNCE_v1-3_preprocessed.zip (~250 MB) using the following command:

gdown https://drive.google.com/uc?id=1fo8F4NKgZDH-bPSdVU3cONAkt5EW-tyr

---

Pretrained Weights

Download the required GroundedSAM model weights here and place them under data/grounded_sam/. The expected directory structure for the data/ folder is as follows:

data
├── grounded_sam
│   ├── groundingdino_swint_ogc.pth
│   ├── GroundingDINO_SwinT_OGC.py
│   ├── repvit_sam.pt
│   └── sam_vit_h_4b8939.pth
├── datasets
│   └── R2R_VLNCE_v1-3_preprocessed
│       ├── embeddings.json.gz
│       ├── envdrop
│       │   ├── envdrop_gt.json.gz
│       │   └── envdrop.json.gz
│       ├── joint_train_envdrop
│       │   ├── joint_train_envdrop_gt.json.gz
│       │   └── joint_train_envdrop.gz
│       ├── test
│       │   ├── test.json
│       │   └── test.json.gz
│       ├── train
│       │   ├── train_gt.json.gz
│       │   └── train.json.gz
│       ├── val_seen
│       │   ├── val_seen_gt.json.gz
│       │   └── val_seen.json.gz
│       └── val_unseen
│           ├── val_unseen_gt.json
│           ├── val_unseen_gt.json.gz
│           ├── val_unseen.json
│           └── val_unseen.json.gz
└── scene_datasets
    └── mp3d
        ├── 17DRP5sb8fy
        │   ├── 17DRP5sb8fy.glb
        │   ├── 17DRP5sb8fy.house
        │   ├── 17DRP5sb8fy.navmesh
        │   └── 17DRP5sb8fy_semantic.ply
        ├── 1LXtFkjw3qL
        │   ├── 1LXtFkjw3qL.glb
        │   ├── 1LXtFkjw3qL.house
        │   ├── 1LXtFkjw3qL.navmesh
        │   └── 1LXtFkjw3qL_semantic.ply
        └── ...

---

Evaluation

OpenNav_R2R-CE_100 download

Download from here to get val_unseen.json.gz and val_unseen_gt.json.gz and place them under data/datasets/R2R_VLNCE_v1-3_preprocessed/val_unseen.

Set the following environment variables before running evaluation:

Variable	Description
LA_API_KEY	API key for the Language Action Model
LA_BASE_URL	Base URL for the Language Action Model endpoint
LA_MODEL_NAME	Model name for the Language Action Model
VA_API_KEY	API key for the Vision Action Model
VA_BASE_URL	Base URL for the Vision Action Model endpoint
VA_MODEL_NAME	Model name for the Vision Action Model

Recommended Models: We recommend using gemini-2.5-pro or gpt-4o as the Language Action Model, and Qwen2.5-VL-32B-Instruct as the Vision Action Model.

Note: Some models (e.g., Qwen3-VL, Gemini) return bounding box coordinates in a relative format. If you wish to use a different Vision Action Model, the coordinate parsing logic may need to be adjusted accordingly.

Configure CUDA_VISIBLE_DEVICES and the number of parallel processes (nprocess) in eval_scripts/r2r.sh, then run:

bash eval_scripts/r2r.sh

On a server equipped with an NVIDIA RTX 4090 GPU using 20 parallel processes, evaluation on the OpenNav-100 episodes completes in approximately 30 minutes.

Simple Visualization

Set inlavira_main.py Line 463：

self.visualize = True

After finishing evaluation, run

python server.py

The website will run on http://0.0.0.0:9999 by default.

Credits

If you find this work useful, please cite our paper:

@article{ding2025lavira,
  title={LaViRA: Language-Vision-Robot Actions Translation for Zero-Shot Vision Language Navigation in Continuous Environments},
  author={Ding, Hongyu and Xu, Ziming and Fang, Yudong and Wu, You and Chen, Zixuan and Shi, Jieqi and Huo, Jing and Zhang, Yifan and Gao, Yang},
  journal={arXiv preprint arXiv:2510.19655},
  year={2025}
}

Our code is adapted from CA-Nav, thanks for their work!