A team led by University of Toronto researchers has used an AI system to turn a chemistry goal into a robot-run laboratory workflow and carry out one quantum-dot experiment with physical equipment. The team’s October 16 preprint describes software that builds a digital model of the available lab, develops the missing robot operations and runs a planned experiment after a researcher approves it. The demonstration automates a particular setup; it does not show an AI independently choosing scientific questions or configuring any laboratory without human input.
How SBL prepares a lab for a new experiment
Most self-driving laboratories automate experiments they were already equipped to run. A new procedure can require engineers to connect instruments, write control software and adapt robot skills. The paper’s “self-building laboratory” system, or SBL, is designed to do some of that preparation for a new task.
A researcher gives SBL a scientific goal and information about the equipment, such as room photographs, device identifiers and available control interfaces. The system builds a digital twin: a computer model of the instruments, their positions and the lab layout. It then uses that model to identify missing capabilities, develop robot skills and assemble an experimental workflow. Sensor feedback from physical trials can help refine precise movements, and successful routines can be saved as scripts for reuse.
The authors describe separate scientific, digital-twin and safety agents. Before physical execution, the system checks candidate operations in simulation. Its physical safeguards include speed limits, force sensing that can trigger a stop and a separate control for a human operator to halt the workflow. The researchers still set the goal, approve the proposed plan and resolve some assumptions the system cannot settle on its own.
One real experiment and wider simulated setups
For its physical demonstration, SBL coordinated liquid handling, a robotic arm, a plate sealer and an optical reader to make cadmium-sulfide quantum dots, tiny semiconductor particles that emit light. It issued 236 liquid-handling commands while moving, sealing and reloading a 96-well plate. The paper reports that the light emitted from one well shifted from 520 to 560 nanometers during an eight-hour reaction.
The evidence has clear boundaries. This single preprint reports a physical quantum-dot workflow, but no controlled comparison showing how much time or engineering effort it saves over existing automation. The authors examined coordination across several robots and workstations in simulation; that broader setup was not part of the physical experiment. They also note that the system depends on what equipment can report and sense: some device failures may not produce an error, and experts were needed to judge physical conditions the software could not observe.
The work suggests a path for lab automation to adapt its software and routines around new goals, rather than relying only on procedures installed in advance. Whether that flexibility generalizes across experiments and reduces the work needed from researchers remains unmeasured.
Sources
Image: Figure 5 from Bai et al., “Self-Building Labs,” licensed CC BY 4.0 via arXiv.