princeton-nlp/WebShop

Provides a simulated e-commerce environment with 1.18M real products and 12K crowdsourced instructions, exposing challenges like compositional reasoning, query reformulation, and strategic exploration. Implements dual rendering modes—HTML for realistic web interaction via browser and OpenAI Gym-compatible text mode for efficient agent training—with baseline models spanning rule-based, imitation learning, and reinforcement learning approaches. Supports sim-to-real transfer evaluation and includes a public Hugging Face demo for interactive testing on live e-commerce sites.

507 stars. No commits in the last 6 months.