LondonAutonomous labs have largely solved robotic movement; experiment-selection models are now the constraint.
An npj Robotics review found actuation sufficient while closed-loop decision models remain unresolved.
Reproducibility audits and cross-lab generalizability are the unsolved limits on autonomous-lab deployment.
Sources: Nature
WHAT A SELF-DRIVING LAB IS
A self-driving lab couples robotic hardware (liquid handlers, synthesis stations, characterization tools) with a decision model that picks the next experiment from the last one's result. The loop runs without a human in it — propose, execute, measure, update, propose again.
THE EASY HALF
Actuation — moving liquids, weighing solids, running an HPLC — is largely solved. Off-the-shelf robotic arms, peristaltic pumps, and OT-2 style liquid handlers reach sub-microliter precision. The mechanical layer is a commodity problem; vendors compete on price, not capability.
THE HARD HALF
Choosing WHICH experiment to run next is the bottleneck. Chemical space is combinatorially vast — a five-component reaction with ten levels each is 100,000 conditions. Bayesian optimization, active learning, and Gaussian processes all work in toy spaces but degrade when the response surface is non-smooth, multimodal, or sparse — which most real chemistry is.
WHY GENERALIZATION FAILS
A model trained on one lab's data rarely transfers to another. Reagent lot, ambient humidity, stirrer geometry, even pipette tip brand shift outcomes enough that a successful campaign in Toronto fails to reproduce in Berlin. The robot is identical; the chemistry isn't.
THE REPRODUCIBILITY AUDIT
Self-driving labs were sold as a fix for chemistry's reproducibility crisis — same code, same hardware, same answer. But the audit trail an autonomous run produces is only as good as the metadata it logs, and most platforms log the moves they made, not the conditions they couldn't measure.
THE COMPARISON TO AI
The arc mirrors self-driving cars: perception and actuation matured fast, but the planning layer — what to do in a novel situation — remains the unsolved core. In both fields, the demo is convincing on the trained distribution and brittle off it.