vlfm

VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation

Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, Bernadette Bucher




Project Website , Paper (arXiv)

:sparkles: Overview

Understanding how humans leverage semantic knowledge to navigate unfamiliar environments and decide where to explore next is pivotal for developing robots capable of human-like search behaviors. We introduce a zero-shot navigation approach, Vision-Language Frontier Maps (VLFM), which is inspired by human reasoning and designed to navigate towards unseen semantic objects in novel environments. VLFM builds occupancy maps from depth observations to identify frontiers, and leverages RGB observations and a pre-trained vision-language model to generate a language-grounded value map. VLFM then uses this map to identify the most promising frontier to explore for finding an instance of a given target object category. We evaluate VLFM in photo-realistic environments from the Gibson, Habitat-Matterport 3D (HM3D), and Matterport 3D (MP3D) datasets within the Habitat simulator. Remarkably, VLFM achieves state-of-the-art results on all three datasets as measured by success weighted by path length (SPL) for the Object Goal Navigation task. Furthermore, we show that VLFM's zero-shot nature enables it to be readily deployed on real-world robots such as the Boston Dynamics Spot mobile manipulation platform. We deploy VLFM on Spot and demonstrate its capability to efficiently navigate to target objects within an office building in the real world, without any prior knowledge of the environment. The accomplishments of VLFM underscore the promising potential of vision-language models in advancing the field of semantic navigation.

:hammer_and_wrench: Installation

Getting Started

Create the conda environment:

conda_env_name=vlfm
conda create -n $conda_env_name python=3.9 -y
conda activate $conda_env_name
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install git+https://github.com/IDEA-Research/GroundingDINO.git@eeba084341aaa454ce13cb32fa7fd9282fc73a67 salesforce-lavis==1.0.2

If you are using habitat and are doing simulation experiments, install this repo into your env with the following:

pip install -e .[habitat]

If you are using the Spot robot, install this repo into your env with the following:

pip install -e .[reality]

[Whether you're using conda or not]

Clone the following repo within this one (simply cloning will suffice):

git clone [email protected]:WongKinYiu/yolov7.git

:dart: Downloading the HM3D dataset

Matterport

First, set the following variables during installation (don't need to put in .bashrc):

MATTERPORT_TOKEN_ID=
MATTERPORT_TOKEN_SECRET=
DATA_DIR=

# Link to the HM3D ObjectNav episodes dataset, listed here:
# https://github.com/facebookresearch/habitat-lab/blob/main/DATASETS.md#task-datasets
# From the above page, locate the link to the HM3D ObjectNav dataset.
# Verify that it is the same as the next two lines.
HM3D_OBJECTNAV=https://dl.fbaipublicfiles.com/habitat/data/datasets/objectnav/hm3d/v1/objectnav_hm3d_v1.zip

Clone and install habitat-lab, then download datasets

Ensure that the correct conda environment is activated!!

…

:weight_lifting: Downloading weights for various models

The weights for MobileSAM, GroundingDINO, and PointNav must be saved to the data/ directory. The weights can be downloaded from the following links:

• mobile_sam.pt: https://github.com/ChaoningZhang/MobileSAM
• groundingdino_swint_ogc.pth: https://github.com/IDEA-Research/GroundingDINO
• yolov7-e6e.pt: https://github.com/WongKinYiu/yolov7
• pointnav_weights.pth: included inside the data subdirectory

:arrow_forward: Evaluation within Habitat

To run evaluation, various models must be loaded in the background first. This only needs to be done once by running the following command:

./scripts/launch_vlm_servers.sh

(You may need to run chmod +x on this file first.) This command will create a tmux session that will start loading the various models used for VLFM and serving them through flask. When you are done, be sure to kill the tmux session to free up your GPU.

Run the following to evaluate on the HM3D dataset:

python -m vlfm.run

To evaluate on MP3D, run the following:

python -m vlfm.run habitat.dataset.data_path=data/datasets/objectnav/mp3d/val/val.json.gz

:newspaper: License

VLFM is released under the MIT License. This code was produced as part of Naoki Yokoyama's internship at the Boston Dynamics AI Institute in Summer 2023 and is provided "as is" without active maintenance. For questions, please contact Naoki Yokoyama or Jiuguang Wang.

:black_nib: Citation

If you use VLFM in your research, please use the following BibTeX entry.

@inproceedings{yokoyama2024vlfm,
  title={VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation},
  author={Naoki Yokoyama and Sehoon Ha and Dhruv Batra and Jiuguang Wang and Bernadette Bucher},
  booktitle={International Conference on Robotics and Automation (ICRA)},
  year={2024}
}