🌎💪 BrowserGym: a Gym Environment for Web Task Automation

🛠️ Setup - 🏋 Usage - 💻 Demo - 🌐 Ecosystem - 🌟 Contributors - 📝 Citation

pip install browsergym

Example of a GPT4-V agent executing openended tasks (top row, chat interactive), as well as WebArena and WorkArena tasks (bottom row).

4x4.grid.mp4

BrowserGym includes the following benchmarks by default:

• MiniWoB
• WebArena
• VisualWebArena
• WorkArena
• AssistantBench
• WebLINX (static benchmark)

Designing new web benchmarks with BrowserGym is easy, and simply requires to inherit the AbstractBrowserTask class.

🛠️ Setup

To use browsergym, install one of the following packages:

pip install browsergym  # (recommended) everything below
pip install browsergym-experiments  # experiment utilities (agent, loop, benchmarks) + everything below
pip install browsergym-core  # core functionalities only (no benchmark, just the openended task)
pip install browsergym-miniwob  # core + miniwob
pip install browsergym-webarena  # core + webarena
pip install browsergym-visualwebarena  # core + visualwebarena
pip install browsergym-workarena  # core + workarena
pip install browsergym-assistantbench  # core + assistantbench
pip install weblinx-browsergym  # core + weblinx

Then setup playwright by running

playwright install chromium

Finally, each benchmark comes with its own specific setup that requires to follow additional steps.

• for MiniWoB++, see miniwob/README.md
• for WebArena, see webarena/README.md
• for VisualWebArena, see visualwebarena/README.md
• for WorkArena, see WorkArena
• for AssistantBench, see assistantbench/README.md

🏗️ Development setup

To install browsergym locally for development, use the following commands:

git clone [email protected]:ServiceNow/BrowserGym.git
cd BrowserGym
make install

Contributions are welcome! 😊

🏋 Usage

Boilerplate code to run an agent on an interactive, open-ended task:

import gymnasium as gym
import browsergym.core  # register the openended task as a gym environment

# start an openended environment
env = gym.make(
    "browsergym/openended",
    task_kwargs={"start_url": "https://www.google.com/"},  # starting URL
    wait_for_user_message=True,  # wait for a user message after each agent message sent to the chat
)
# run the environment <> agent loop until termination
obs, info = env.reset()
while True:
    action = ...  # implement your agent here
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        break
# release the environment
env.close()

MiniWoB

import gymnasium as gym
import browsergym.miniwob  # register miniwob tasks as gym environments

# start a miniwob task
env = gym.make("browsergym/miniwob.choose-list")
...

# list all the available miniwob tasks
env_ids = [id for id in gym.envs.registry.keys() if id.startswith("browsergym/miniwob")]
print("\n".join(env_ids))

WorkArena

import gymnasium as gym
import browsergym.workarena  # register workarena tasks as gym environments

# start a workarena task
env = gym.make("browsergym/workarena.servicenow.order-ipad-pro")
...

# list all the available workarena tasks
env_ids = [id for id in gym.envs.registry.keys() if id.startswith("browsergym/workarena")]
print("\n".join(env_ids))

WebArena

import gymnasium as gym
import browsergym.webarena  # register webarena tasks as gym environments

# start a webarena task
env = gym.make("browsergym/webarena.310")
...

# list all the available webarena tasks
env_ids = [id for id in gym.envs.registry.keys() if id.startswith("browsergym/webarena")]
print("\n".join(env_ids))

VisualWebArena

import gymnasium as gym
import browsergym.webarena  # register webarena tasks as gym environments

# start a visualwebarena task
env = gym.make("browsergym/visualwebarena.721")
...

# list all the available visualwebarena tasks
env_ids = [id for id in gym.envs.registry.keys() if id.startswith("browsergym/visualwebarena")]
print("\n".join(env_ids))

AssistantBench

import gymnasium as gym
import browsergym.workarena  # register assistantbench tasks as gym environments

# start an assistantbench task
env = gym.make("browsergym/assistantbench.validation.3")
...

# list all the available assistantbench tasks
env_ids = [id for id in gym.envs.registry.keys() if id.startswith("browsergym/workarena")]
print("\n".join(env_ids))

💻 Demo

If you want to experiment with a demo agent in BrowserGym, follow these steps

# conda setup
conda env create -f demo_agent/environment.yml
conda activate demo_agent

# or pip setup
pip install -r demo_agent/requirements.txt

# then download the browser for playwright
playwright install chromium

Our demo agent uses openai as a backend, be sure to set your OPENAI_API_KEY.

Launch the demo agent as follows

# openended (interactive chat mode)
python demo_agent/run_demo.py --task_name openended --start_url https://www.google.com

# miniwob
python demo_agent/run_demo.py --task_name miniwob.click-test

# workarena
python demo_agent/run_demo.py --task_name workarena.servicenow.order-standard-laptop

# webarena
python demo_agent/run_demo.py --task_name webarena.4

# visualwebarena
python demo_agent/run_demo.py --task_name visualwebarena.398

You can customize your experience by changing the model_name to your preferred LLM (it uses gpt-4o-mini by default), adding screenshots for your VLMs with use_screenshot, and much more!

python demo_agent/run_demo.py --help

🌐 Ecosystem

• AgentLab: Seamlessly run agents on benchmarks, collect and analyse traces.
• WorkArena(++): A benchmark for web agents on the ServiceNow platform.
• WebArena: A benchmark of realistic web tasks on self-hosted domains.
• VisualWebArena: A benchmark of realistic visual web tasks on self-hosted domains.
• MiniWoB(++): A collection of over 100 web tasks on synthetic web pages.
• WebLINX: A dataset of real-world web interaction traces.
• AssistantBench: A benchmark of realistic and time-consuming tasks on the open web.

🌟 Contributors

📝 Citing This Work

Please use the following BibTeX to cite our work:

@inproceedings{workarena2024,
    title = {{W}ork{A}rena: How Capable are Web Agents at Solving Common Knowledge Work Tasks?},
    author = {Drouin, Alexandre and Gasse, Maxime and Caccia, Massimo and Laradji, Issam H. and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre},
    booktitle = {Proceedings of the 41st International Conference on Machine Learning},
    pages = {11642--11662},
    year = {2024},
    editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix},
    volume = {235},
    series = {Proceedings of Machine Learning Research},
    month = {21--27 Jul},
    publisher = {PMLR},
    url = {https://proceedings.mlr.press/v235/drouin24a.html},
}