Awesome AI Agents › Web Automation Systems

stanfordnlp/wge

⭐ 118 Python added to this list on 2025-04-19 repository created 2018-02-20

The project "wge" (Workflow-Guided Exploration) is a research codebase developed by Stanford NLP group, designed to train reinforcement learning (RL) agents to perform tasks on web interfaces based on natural language instructions. The core idea is to enable machine learning models to interact with web browsers to complete tasks such as booking flights or filling out forms by understanding and executing commands given in natural language. This project accompanies a research paper presented at ICLR 2018 titled "Reinforcement Learning on Web Interfaces using Workflow-Guided Exploration." The codebase supports reproducible experiments and provides detailed setup instructions for dependencies including Python packages, Node.js, PyTorch, and Selenium for browser automation. The system uses a combination of environments, policies, episode generators, and replay buffers to train and evaluate RL agents. The environment simulates web tasks, with MiniWoB (Mini Web Benchmark) and FormWoB being key environments used. Policies represent the agents' decision-making models, including both the main model policy and an exploration policy. Episode generators produce sequences of actions (episodes) by running policies in environments, and replay buffers store these episodes to train the model policy effectively. The project supports batch processing to speed up training by running multiple environments in parallel. It also provides a modular architecture where different components like policies and environments implement specific interfaces, allowing extensibility and experimentation. Configuration is managed through HOCON format files, enabling flexible experiment setups. The repository includes scripts to launch training experiments, manage training runs, and evaluate models. Overall, this project is a sophisticated framework for advancing research in applying reinforcement learning to web-based tasks, focusing on sample efficiency and workflow-guided exploration strategies. It is valuable for researchers and developers interested in natural language interfaces, RL, and web automation.

https://github.com/stanfordnlp/wge

browser-automationenvironmentepisode-generatorformwobiclr-2018machine-learningminiwobnatural-language-instructionspolicypytorchreinforcement-learningreplay-bufferrl-agentsample-efficiencyseleniumtrainingweb-tasksworkflow-guided-exploration