A fake web store stocked with 1.18 million real products

What it does WebShop is a research environment that simulates an online store using 1.18 million real-world products and 12,087 crowd-sourced instructions. An agent receives a natural language goal—say, finding a specific pair of shoes—and must search, click through results, and handle item attributes until it makes a purchase. The project ships with both a full browser-based interface and a stripped-down text mode that exposes a standard OpenAI Gym API.

The interesting bit The environment straddles the line between toy gridworlds and the actual internet: it serves real HTML pages locally, but packages them into a clean reinforcement-learning interface with reward signals. The authors also provide a Hugging Face demo that attempts sim-to-real transfer, letting agents trained in the simulator loose on live Amazon and eBay listings.

Key highlights

• 1.18 million real products and 12,087 human-written instructions provide a large, grounded action space.
• Two observation modes: a rendered html mode with clickable pages and a simple text mode for faster model training.
• Includes baseline implementations covering rule-based, imitation learning, RL, and hybrid IL+RL agents.
• Provides trajectory logging and attribute inspection flags for analyzing agent behavior.
• Sim-to-real transfer experiments are supported via a separate transfer pipeline and an online demo.

Caveats

• By default the environment loads only 1,000 products; using the full catalog requires manually editing paths in web_agent_site/utils.py.
• The full browser mode requires Java and ChromeDriver, making headless server deployment more involved than a pure Python package.
• The README notes the project is primarily for reproducing paper results, so the codebase is research-grade rather than a polished SDK.

Verdict Researchers building grounded language agents or studying web navigation and sim-to-real transfer should take a look. If you just need a web-scraping tool or a drop-in RL environment without web-page semantics, this is over-engineered for your needs.