Self-Building Labs

Self-Building Labs presents a system that turns a scientific goal and a description of available equipment into a working laboratory capability. Rather than requiring specialists to manually reconfigure each self-driving lab, its agents identify missing capabilities, construct a digital twin, develop robot skills and experimental workflows, then validate them in simulation and through physical feedback. The implementation combines OpenAI’s GPT-6-Astra with NVIDIA Isaac Sim and cuRobo; successful exploratory actions are distilled into reusable scripts, while safety agents test hazards in simulation and runtime guardrails constrain physical commands. In a real quantum-dot experiment, the system used an Opentrons OT-2 to prepare a 96-well plate, then coordinated robotic transfers for sealing and optical measurement: the protocol issued 236 commands, followed by 19 robotic stages. The plate was monitored at 60 °C for 8 h, and the measurements showed stronger quantum-dot emission under high-pH conditions, with ligand loading affecting formation. The paper also demonstrates coordination across different robot types and workstations in simulation, suggesting a route from fixed, task-specific automation toward laboratories that can build and reuse capabilities as research needs change.