AI agents design and run a robot chemistry lab in University of Toronto demo
A preprint describes “self-building labs” that add missing robot skills in a digital twin, then try them on real instruments.
Researchers led by the University of Toronto’s Acceleration Consortium describe AI agents that can design, configure and operate a robot-powered chemistry lab from a high-level scientific goal, in an arXiv preprint posted Oct. 30, 2026.
Ordinary “self-driving labs” automate experiments they were already wired for. The authors’ “self-building lab” starts from whatever is in the room—even an empty space—and fills gaps: a digital twin of the equipment, robot skills and step-by-step workflows, tested in simulation before running on real machines.
In one demo, the system wrote more than 200 commands for an Opentrons liquid-handling robot, then used a rail-mounted arm to seal a plate and load it into an optical reader to track cadmium sulfide quantum-dot growth, the paper said. NVIDIA Isaac Sim powered the twin; OpenAI’s GPT-6 Astra helped plan.
The work is a preprint and has not been peer-reviewed. Human operators still approve plans and handle stock chemicals.