Autonomous Generalist Scientist: Towards and Beyond Human-level Automatic Research Using Foundation Model-Based AI Agents and Robots.

The accelerating pace of scientific research highlights the need for more efficient, accurate, and comprehensive methodologies. Traditional research methods are often hindered by the limitations of manual experimentation and data collection in isolated environments, resulting in slow and resource-intensive processes. Besides, multidisciplinary research presents significant challenges due to the complexities of integrating knowledge from various fields, often surpassing the expertise of individual researchers. The limited knowledge base of single researchers constrains the scope and depth of inquiry, complicating efforts to fully explore complex interdisciplinary problems. To address these challenges, it is crucial to develop automatic systems that can dynamically interact with both physical and virtual environments while facilitating the integration of knowledge across multiple disciplines. Foundation AI models, such as large language models, are trained on vast amount of data from diverse sources, enabling them to acquire knowledge across various scientific disciplines. Therefore, it is promising to build the generalist AI robot scientist for autonomous research based on these foundation models and robot technologies.

Auto Research Agents framework and vision

The Autonomous Generalist Scientist （AI Scientist） project aims to revolutionize the academic research process by introducing a framework for fully automated research agents/robots. This initiative seeks to integrate artificial intelligence into every stage of research—from literature reviews to proposal, experiment, writing, submitting, reviewing manuscripts. Our vision is to facilitate a seamless research workflow that enhances productivity and fosters innovation in scientific inquiry. Here’s how we envision the integration of automated agents and robotics evolving within the framework:

Phase 1: Software-Only Agents Initially, the project will focus on software-only agents that can perform tasks not requiring physical interaction with the real world.

Phase 2: Integration of Robotics As the project matures and the capabilities of our agents evolve, we plan to introduce robotics to carry out physical tasks and experiments in the laboratory.

Auto Research Timeline

• gscientist/: The main directory housing all the components of Nova.
  • core/: System management.
  • agents/: Contains the agent implementations with agent.py serving as a template or a particular instance.
  • communication/: Contains the ros like communication mechanism.
  • llm/: Dedicated to language model functionality, where llm.py defines language model-related operations.
  • tools/: A suite of utility scripts to augment the functionality of agents.
    • builtins/: Essential tools that come with the Nova framework.
    • plugins/: Dynamically loaded modules that extend the capabilities of Nova agents.
• ui/: User interface assets that define how users interact with the Nova framework.
  • frontend/: Web ui.
  • qt/: The graphical user interface components built with PyQt or PySide6 for user interaction.

Todo: Instructions on how to set up the environment, install dependencies, and run the initial configuration.

To make a contribution or suggest an idea to this project.

• Discord server

• Outline-Google doc

bibtex
@misc{auto-research,
  title        = {Autonomous Generalist Scientist: Towards and Beyond Human-level Automatic Research Using Foundation Model-Based AI Agents and Robots (A Position)},
  author       = {Starkson Zhang, Alfredo Pearson, Zhenting Wang, Welcome Coauthors},
  year         = {2024},
  journal      = {DOI: 10.13140/RG.2.2.35148.01923},
}