Closed loop experiments at Berkeley's A Lab, a self driving materials lab, show that 24/7 AI robot systems are moving from curiosity into the physical infrastructure of materials research, and turning reproducibility into a design problem.
In the Hearst Memorial Mining Building at UC Berkeley, just past midnight, a row of sake-cup-sized crucibles slides out of an industrial oven. Inside each one is a powder blend of metal oxides, the precursor for a new battery or catalyst that may or may not work. The crucibles move into a centrifuge, are mixed with zirconium beads, then sent for x-ray diffraction and ionic-conductivity measurements. The result of every step is fed to a planning agent, which proposes the next experiment while no one is in the room. By morning, a researcher walking in will find a queue of overnight runs already finished, and a list of next attempts ready to load.
This is the A-Lab at Lawrence Berkeley National Laboratory, a "self-driving" lab that pairs robots and AI to run materials and chemistry experiments continuously, around the clock, with a human supervising from a Slack channel rather than standing at the bench. The Scientific American feature by Patrick Sisson describes a working version of something materials researchers have been promising for nearly a decade: a closed loop where an AI proposes the next experiment, a robot executes it, the instrument characterizes the result, and the result feeds the next proposal, often within minutes, and often through the small hours when a human is asleep or off shift.
The category has a name, "lab in the loop," and a useful cousin. In a human-in-the-loop lab, a grad student actively drives the experiment: choosing the next sample, fixing parameters, deciding when to stop. In a human-on-the-loop lab, the AI runs, the human supervises, and the human only steps in when something breaks. The A-Lab lives in the second mode, and that distinction is the news. Materials science has always been a slow, hands-on discipline, paced by graduate students who can run a handful of reactions a day. A closed-loop robot system can run dozens, sometimes hundreds, with the planning software tightening the search each round.
The team at Berkeley is led by Gerbrand Ceder, one of the most cited computational materials scientists in the field, and the agents that drive the lab are a small cast with names: Minerva, Alfred, Prometheus, and Jeeves, drawn from mythology and literature because, as the piece notes, the researchers wanted the machines to feel like members of the lab rather than anonymous black boxes. Each has a job. One plans the next experiment from the prior result. One watches for faults. One manages the queue of precursors and samples. One keeps the documentation. The point is not the naming. The point is the division of cognitive labor, and the way the system keeps going when the hardware misbehaves. Sisson reports that when a rack jams or a sample spills, the agents replan on the fly and route around the failed step rather than halting the run.
This is what makes the A-Lab feel like infrastructure rather than a one-off stunt. A single closed-loop demonstration is a paper. A system that detects its own faults, replans, and continues the experiment is a piece of working scientific infrastructure, with all the maintenance, calibration, and trust issues that come with infrastructure. The same article puts the open question bluntly: the trust question is unresolved. The robots do not get tired, but they do drift. Instruments go out of calibration. AI planning models can hallucinate plausible-looking experiments that no real chemistry would produce. A run that finishes 200 attempts overnight is only valuable if the results are reproducible, and reproducible, in this setting, means a careful human-readable paper trail, version-controlled prompts, and characterization instruments that someone checks.
Those design problems are now the story. The legitimate critique of self-driving labs is not that they are scary; it is that the reliability of a single overnight run is the average of every calibration, every prompt update, every batch of precursor. Reproducibility stops being a footnote and becomes a front-end design choice. Sisson quotes Ceder asking, "Can you build an AI that acts like a scientist?" The harder question, the one that will define whether this generation of A-Lab successors becomes standard equipment in a chemistry department, is whether the field can build the calibration, documentation, and shared standards that make those overnight runs trustworthy enough to publish.
The use cases already stretch beyond batteries. The piece points to autonomous systems aimed at catalysis and at early-stage cancer-therapy discovery, where the closed loop is the same: a planning agent proposes, a robot mixes and measures, a result feeds the next iteration. If even a fraction of those efforts deliver, the bottleneck in early-stage materials and chemistry research shifts from the time a human can stay in the lab to the time a model can be trusted to run unsupervised.
What to watch next is concrete. The first signal that self-driving labs are becoming real infrastructure, rather than Berkeley's signature project, will be the day an A-Lab result is published with a machine-readable methods section detailed enough for a competing lab to replay it overnight. Until then, the 24/7 lab is a working proof of concept, and the interesting work is the unglamorous, essential work of giving the robots a paper trail a human can defend.