WebShop provides a simulated e-commerce environment for training and evaluating grounded language agents. It addresses challenges in real-world web interaction, such as understanding complex instructions, reformulating queries, and handling noisy web content. The environment is designed for researchers and developers in NLP and reinforcement learning focused on building agents capable of complex web navigation and task completion.

How It Works

WebShop simulates an e-commerce website with over a million real products and crowd-sourced instructions. Agents interact with the environment by issuing commands like search, click, and select. The environment offers two observation modes: 'html' for a richer, browser-like experience and 'text' for a simplified, OpenAI Gym-compatible interface. This dual approach allows for both realistic interaction modeling and streamlined agent development.

Quick Start & Requirements

• Install: Clone the repository and run ./setup.sh [-d small|all].
• Prerequisites: Python 3.8.13, Java.
• Optional: ResNet image features, Human demonstration data, ChromeDriver (for run_web_agent_site_env.sh).
• Data: The setup.sh script downloads product/instruction data and a spaCy model. The full dataset can be enabled by modifying web_agent_site/utils.py.
• Demo: A Hugging Face demo is available.

Highlighted Details

• Simulated e-commerce environment with 1.18M products and 12,087 instructions.
• Supports both HTML and simplified text observation modes.
• Includes baseline models (rule, IL, RL) and sim-to-real transfer code.
• Hugging Face demo available for interactive exploration.

Maintenance & Community

The project is from Princeton NLP. Contributions are welcomed via pull requests and issues.

Licensing & Compatibility

The license is available in LICENSE.md. No specific license type is mentioned in the README, but it is generally expected to be permissive for research.

Limitations & Caveats

The run_web_agent_site_env.sh script requires a specific ChromeDriver version compatible with the user's Chrome browser. The README does not explicitly state the license type, which may impact commercial use.