Toronto team introduces self-building labs with AI digital twins for chemistry
Researchers from the University of Toronto, Vector Institute, Acceleration Consortium, Canadian Institute for Advanced Research and NVIDIA have introduced a self-building lab framework that uses AI agents, digital twins and robots to expand laboratory automation beyond fixed workflows. In an arXiv paper titled Self-Building Labs, co-first authors Jiaru Bai and Yuchi Zhao, with co-corresponding authors Yang Cao, Varinia Bernales and Alan Aspuru-Guzik, describe systems that can interpret scientific goals, inventory available equipment, identify capability gaps and help create the interfaces, skills and procedures needed to carry out experiments. The framework is distinct from self-driving labs, which automate execution within hardware and software limits established in advance. Self-building labs instead aim to move the definition and extension of laboratory capabilities into the automation loop. The approach relies on a digital twin of the physical lab, constructed from photos, device models, manufacturer information and robotic perception. Three agents cooperate: a Scientific Agent converts research questions into experimental protocols; a Twin Agent translates those protocols into robot commands and develops missing device interfaces or motion skills; and a Safety Agent explores hazardous scenarios in simulation and generates safety rules. The agents share logs, simulation results and feedback from real experiments to improve future performance. The team tested the method by synthesizing cadmium sulfide quantum dots. The system generated and validated a 236-step Python protocol for an OT-2 liquid handler, managing device initialization, pipetting speed, tip handling and reagent preparation. Robots then coordinated transfer of a 96-well plate to a sealing instrument, re-gripped the plate and loaded it into a Cytation 5 detector for optical measurement, completing 19 coordinated action stages. Because simulation errors can affect fine manipulations, the robot used vision, depth and force-torque feedback during critical operations. Later demonstrations extended the approach to flow photochemistry, online nuclear magnetic resonance analysis and high-performance liquid chromatography, including drawer operation and plate transfer between stations. The researchers also envision broader uses such as multi-lab coordination, robotic sample transport across facilities and eventual design or fabrication of custom lab hardware through additive manufacturing. The team emphasizes that humans remain essential for setting boundaries, identifying risks and providing physical observations that instruments may miss. The system does not replace scientific judgment or predictive chemistry; it focuses on robotic manipulation, device integration and workflow expansion. The work suggests laboratory automation may shift from forcing scientists to adapt to existing platforms toward enabling labs to adapt to new scientific questions.
