Awesome AI Agents › Browser Automation Tools

posgnu/rci-agent

⭐ 241 HTML added to this list on 2025-04-19 repository created 2023-04-02

The project "RCI Agent for MiniWoB++" presents a codebase implementing an agent that leverages pre-trained language models to solve computer tasks within the MiniWoB++ benchmark environment. This agent uses a novel RCI prompting scheme to enhance its task execution capabilities guided by natural language instructions. The MiniWoB++ benchmark is a suite of web-based tasks designed to evaluate the ability of AI agents to interact with web environments effectively. The RCI agent is implemented in Python and integrates with the OpenAI Gym environment, requiring dependencies such as gym, openai, selenium, Pillow, and regex. Users must install MiniWoB++ and configure their OpenAI API key to run the agent. The agent supports various language models, including GPT-3.5-turbo (chatgpt), text-davinci-003, and others, allowing flexibility in experimentation. The project provides detailed instructions for setup, running tasks, and configuring parameters such as the number of episodes, explicit and implicit RCI loops, and state grounding updates. The evaluation results demonstrate that the RCI agent achieves competitive performance, ranking second among tested models while using significantly fewer training samples compared to other approaches like WebN-T5-3B and CC-Net. This efficiency highlights the potential of large language models (LLMs) to solve complex computer tasks without extensive expert demonstrations or reward function engineering. The project is supported by a research paper available on arXiv, which should be cited when using the code. Overall, this project contributes to advancing the use of language models in interactive task-solving environments, showcasing a promising direction for AI research in automating computer-based tasks through natural language understanding and interaction.

https://github.com/posgnu/rci-agent

ai-agentcomputer-tasksevaluationgpt-3.5-turbolanguage-modelslarge-language-modelsllmminiwob++natural-languageopenai-apiopenai-gympre-trained-language-modelpromptingpythonrci-agentreasoningreinforcement-learningsample-efficiencytask-automationtask-executiontext-davinci-003web-environment