World of Bits

World of Bits (WoB): An OpenAI-pioneered reinforcement learning platform that trains AI agents to complete complex, open-domain web tasks using low-level keyboard and mouse actions.

World of Bits (WoB) is a critical reinforcement learning platform, initially developed at OpenAI by researchers like Andrej Karpathy, designed to train AI agents for generalist web interaction. It moves past simulated environments by packaging a Chrome browser within a Docker container, exposing a Gym interface for the agent. The agent receives observations via the Document Object Model (DOM) and rendered pixels, then performs low-level actions (keyboard, mouse clicks) to complete tasks on the live web. This methodology, which includes the smaller, controlled Mini World of Bits (MiniWoB) environment, provides a rich, human-created source of tasks—like interacting with buttons and date pickers—for developing generalist agents capable of navigating the billions of hours humans spend on digital devices.

https://openai.github.io/signup-forms/