BrowserGym Core

0.14.3 · active · verified Sat Apr 11

BrowserGym Core (version 0.14.3) is a Python library that provides the fundamental Gymnasium environment for web task automation within the Chromium browser. It serves as the base for various web agent benchmarks and is designed for researchers and developers evaluating and building web agents, particularly those leveraging Large Language Models (LLMs) for web interaction tasks. The project is under active development with a rapid release cadence.

Warnings

gotcha BrowserGym requires Playwright and its Chromium browser to be installed separately. Missing `playwright install chromium` is a common setup error.
Fix: Run `playwright install chromium` after `pip install browsergym-core`.
gotcha BrowserGym is explicitly stated as a research tool and not a consumer product. Users should be aware it's under active development and may have rough edges.
Fix: Set appropriate expectations for stability and feature completeness; contribute to its development if issues arise.
gotcha Timeout errors are frequently encountered when interacting with web pages. These can be due to slow loading, complex elements, or agent actions that don't complete in time.
Fix: Implement robust error handling, increase timeout parameters where applicable, and refine agent logic for complex web interactions. Consider debugging with `headless=False` to observe browser behavior.
gotcha When integrating `browsergym-core` with specific benchmarks (e.g., WebArena), additional environment variables (e.g., `WA_PORT`) or benchmark-specific setup steps are often required to correctly configure task URLs and instances.
Fix: Consult the `README.md` or documentation for the specific `browsergym-*` benchmark package you are using for its unique setup instructions.
breaking Following Gymnasium API changes, `env.reset()` now returns `(observation, info)` and `env.step()` returns `(observation, reward, terminated, truncated, info)`. Older `(observation, info)` from `reset()` or `(observation, reward, done, info)` from `step()` patterns are deprecated.
Fix: Update your code to match the `gymnasium` API: `obs, info = env.reset()` and `obs, reward, terminated, truncated, info = env.step(action)`.

Install

pip install browsergym-core Install core library
playwright install chromium Install Playwright browser

Imports

gym
```
import gymnasium as gym
```

BrowserEnv

from browsergym.core.env import BrowserEnv

AbstractBrowserTask

from browsergym.core.task import AbstractBrowserTask

Quickstart

This quickstart demonstrates how to initialize a `browsergym-core` environment for an open-ended task, perform a reset, take a sample action, and ensure proper cleanup. It uses `gymnasium` for the environment API and `playwright.sync_api` to manage the browser context.

import gymnasium as gym
from playwright.sync_api import sync_playwright

# Register the 'openended' task provided by browsergym-core
import browsergym.core

# It's good practice to install playwright chromium if not already done:
# playwright install chromium

with sync_playwright() as p:
    # Initialize the BrowserGym environment for an open-ended task
    # 'headless=True' runs the browser without a visible UI
    env = gym.make(
        "browsergym/openended", 
        headless=True,
        task_kwargs={'start_url': 'about:blank'}
    )

    try:
        obs, info = env.reset()
        print(f"Initial observation keys: {obs.keys()}")
        print(f"Initial info: {info}")

        # Example: taking a no-op step
        action = env.action_space.sample()
        obs, reward, terminated, truncated, info = env.step(action)

        print(f"Reward after one step: {reward}")
        print(f"Terminated: {terminated}, Truncated: {truncated}")

    finally:
        # Ensure the browser environment is closed properly
        env.close()

view raw JSON →