create a conda environment and install requirements.
conda create -n virl python=3.8
conda activate virl
pip install -r requirement.txt
Build the package
python setup.py develop
V-IRL relies on Google Map Platform
export GOOGLE_MAP_API_KEY="YOUR_API_KEY"
export OPENAI_API_KEY="YOUR_API_KEY"
export SERPAPI_API_KEY="YOUR_API_KEY" # optional, used in estate agent vivek only
If you don't have the API key for Google MAP Platform, please visit Google Map Platform to get one. You would have $200
credits monthly, which is sufficient for most V-IRL
exemplar agents.
Please visit rentcast to get the api key.
In V-IRL
, we use selenium to control the web browser to interact with the web page using python.
In current version, you should NOT need to explicitly install Selenium. However, if the automatic installation fail, you may need to install it manually as follow:
Take Chrome as an example, you can download suitable driver following this. Notice that you have to check your Chrome version and download the corresponding driver.
If you just want to attempt agents that not requiring vision abilities such as: Peng
, Aria
, Vivek
and Local
, you can just skip this section.
However, if you want to run other agents, you should deploy vision models as follows:
To play V-IRL
agents in any computer (with or without GPU), we decouple the vision model deployment and V-IRL agents/benchmark running.
Important Note: Thanks to our decoupling design, you don't need to install all following vision models in single conda environment and machine.
Step 1: Clone and install our custom GLIP.
Step 2: Launch the GLIP server
python glip_app.py --model_size large
Step 3: Record the IP and port to the configure
GLIP:
ENABLED: True
ROOT_PATH: ../../GLIP
MODEL_SIZE: large # [tiny, large]
# for GLIP CLIENT ONLY
SERVER: http://xxx.xxx.xxx.xxx:xxxx # your ip and port here
Step 1: Launch CLIP server on any machine
cd virl/perception/recognizer
python clip_server.py
Step 2: Record the IP and port to the configure
CLIP:
SERVER: http://xxx.xxx.xxx.xxx:xxxx # your ip and port here
Step 1: Clone and install our custom LightGlue.
Step 2: Launch the LightGlue server
python app.py
Step 3: Record the IP and port to the configure
LIGHTGLUE:
SERVER: http://xxx.xxx.xxx.xxx:xxxx # your ip and port here