PNNL

Abstract

Accelerating scientific discovery through automation is hampered by the critical gap between a researcher’s high-level intent and the low-level code required by robotic systems. To address this challenge, we show that a cognitive artificial intelligence (AI) architecture, comprising collaborative and self-correcting agents, can bridge this gap by autonomously designing complex chemical experimental protocols from natural language interactions. As a demonstration, we develop AutoLabs, a generative AI system for chemical procedure planning and implementation for an automated liquid handling robotic system. We perform systematic benchmarking by developing a series of test cases of increasing complexity, implementing quantitative metrics to probe generated procedure correctness, and performing end-to-end evaluation under different levels of human collaboration. This benchmarking reveals that the AI agent’s underlying reasoning capacity is the most decisive factor for success. Implementing a strong reasoning model alone reduces quantitative errors in stoichiometry by over 85% and enables the system to correctly parse intricate procedural logic. This reasoning core, when augmented within a multi-agent framework with specialized tools and iterative self-correction, achieves high procedural accuracy (F1-score > 0.89) on challenging multi-step syntheses. Our work establishes a design blueprint for creating robust and reliable AI partners for chemical experimentation, providing a pathway towards the next generation of autonomous laboratories. Code https://github.com/pnnl/autolabs.