A robot can automate researcher-facing lab work when search and safety are bounded.
Executive brief
The operating-model hinge is linking experiment choice, physical execution, measurement, and fault handling.
AI value · Autonomous search
Verified688 experiments in eight days; best mixture six times more active
Laboratory proof of concept, not a production deployment; 11 results were discarded.
Before
Chooses formulations, prepares samples, operates instruments, records results, and decides the next experiment. → Interprets results and updates the experimental plan.
After
Selects the next batch from prior measured outcomes. → Navigates the lab, dispenses materials, runs photolysis and chromatography, and returns measurements.
Human boundary
Researchers set the problem and safety envelope; the system selects and executes experiments within it.
Why it matters
A robot can automate researcher-facing lab work when search and safety are bounded.
Before
How the work ran before the change.
Human chemist
Chooses formulations, prepares samples, operates instruments, records results, and decides the next experiment.
ControlLaboratory safety and researcher judgment; authors describe manual search as constrained but provide no measured manual baseline.
Human researcher
Interprets results and updates the experimental plan.
ControlManual trial-and-error across a ten-variable space.
What changed
A robot can automate researcher-facing lab work when search and safety are bounded.
Decision rightAI acts within a human backstop
After
How the same work runs now.
Batched Bayesian search algorithm
Selects the next batch from prior measured outcomes.
ControlTen-variable bounded search; oxygen and workflow checks discard faulty results.
Mobile robot
Navigates the lab, dispenses materials, runs photolysis and chromatography, and returns measurements.
ControlCollaborative-robot safety standards; 11 results discarded for workflow errors or high oxygen.
Exception path
Faulty seals, high oxygen, workflow errors, or hardware issues are flagged/discarded for human attention.
Decision authority
Researchers set the problem and safety envelope; the system selects and executes experiments within it.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Human chemist | Chooses formulations, prepares samples, operates instruments, records results, and decides the next experiment. | Laboratory safety and researcher judgment; authors describe manual search as constrained but provide no measured manual baseline. |
| 02 | Human researcher | Interprets results and updates the experimental plan. | Manual trial-and-error across a ten-variable space. |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Batched Bayesian search algorithm | Selects the next batch from prior measured outcomes. | Ten-variable bounded search; oxygen and workflow checks discard faulty results. |
| 02 | Mobile robot | Navigates the lab, dispenses materials, runs photolysis and chromatography, and returns measurements. | Collaborative-robot safety standards; 11 results discarded for workflow errors or high oxygen. |
Work that left the path
Human role before
Researchers perform repetitive preparation, movement, analysis, and next-experiment selection.
Human role after
Researchers define the objective and supervise a robot that executes and searches continuously.
AI role
Selects experiment batches and controls a mobile robot across laboratory stations.
Initial catalyst formulations688 experiments in eight days; best mixture six times more active
172 operating hours across eight days · Ten-variable photocatalyst search; 319 moves and 6,500 manipulations
Laboratory proof of concept, not a production deployment; 11 results were discarded.
Anti-pattern
Do not generalize one autonomous experiment to unsupervised general chemistry.
Questions
Portability conditions
Reputation risk
high