Environment-Aware Multi-Agent Framework for Self-Driving Laboratories

Most self-driving laboratories (SDLs) rely on the assumption of a static and fully known experimental environment prior to execution, neglecting potential pre-experimental deviations and human errors that may trigger cascading failures during operation.To address these limitations, this paper presents a multi-agent SDL framework incorporating pre-execution environment perception. The system is designed to emulate human-like manipulation intelligence, integrating natural language interaction, synthesis planning, dexterous environment perception, cognitive task planning, and adaptive robot execution into a closed-loop workflow. The system is coordinated through six specialized agents-task clarification, synthesis planning, environment perception, robot planning, robot execution, and feedback-collectively enabling end-to-end automation from user intent to experimental realization. Validation was conducted on a customized automation platform using hydrogel and silicone preparation tasks, supported by a dedicated dataset encompassing these processes and multiple types of potential anomalies inconsistent with predefined conditions. Experimental results demonstrate high accuracy in both synthesis planning and environment perception (with perception accuracy reaching 99.86%), along with significant improvements in task success rates: hydrogel preparation increased from 0.64 to 0.90, and silicone preparation from 0.60 to 0.92. These findings confirm that integrating environment perception with multi-agent collaboration effectively enhances system robustness, safety, and adaptability under unexpected anomalies, underscoring the framework’s potential for scalable deployment in real-world laboratory environments.