A commercial robot closes the lab loop: agentic optimization needs up to 52% fewer iterations than Bayesian optimization
Synopsis
A self-driving-laboratory testbed built from a commercial lab robot, an absorbance plate reader and a swappable optimizer completed six independent campaigns of 48 closed-loop iterations each, over 24 cumulative hours of unattended operation. Five of the six campaigns matched their target absorbance spectra within measurement error, at an average RMSE of OD. In the same physical loop, an agentic optimizer driven by Claude Opus 5 matched all three targets in fewer iterations than Bayesian optimization, cutting iterations by 52% on target 1 and 32% on target 2, and converging in 11 iterations on target 3, which Bayesian optimization failed to match within 48.
Interpretation
Self-driving laboratories now have a physical testbed assembled from off-the-shelf commercial instruments that can run the full design-build-test-learn loop unattended on one hardware setup. Earlier closed-loop platforms were mostly bespoke, single-instance systems aimed at specific scientific problems, and integration of commercial equipment with AI decision-making remained at an early stage, with no shared reference for comparing decision algorithms on a low-risk task. Six independent campaigns of 48 iterations each, about four hours per campaign and more than 24 cumulative hours unattended, with every step from pipetting and plate reading to choosing the next formulation performed by the system.
The testbed gives liquid formulation and spectral measurement to the robot and hands the choice of the next formulation to a swappable decision algorithm. The decision layer can be the open-source Bayesian optimization library BoTorch or a Claude Opus 5 agent called through the Claude Messages API, which writes and runs its own modeling code in a local Python sandbox. The task mixes red, yellow and blue dyes with water under a fixed 250 µL total volume, reads absorbance at 400, 600 and 700 nm, and judges a match against the plate reader's measurement accuracy specifications.
The agent hit all three targets faster than Bayesian optimization and needed no hand-specified optimizer structure. Bayesian optimization required custom BoTorch code, manual hyperparameter selection and a pre-specified model structure, while the agent selected and implemented its own models. Target 1 fell from 21 to 10 iterations and target 2 from 34 to 23, while target 3 was unmatched by Bayesian optimization within 48 iterations and matched by the agent in 11; per-iteration decision time was comparable, averaging 19 seconds for the agent and 16 seconds for Bayesian optimization.
Perspective
The demonstrated scope is low-viscosity liquid dispensing and filter-based absorbance at three fixed wavelengths, on a task that matches target spectra with three dyes and water under a fixed 250 µL total volume; the authors state that the workflow, while representative, does not establish generalizability across the full range of experimental techniques or material systems in chemistry, materials science and biology.
The authors state that a fully agentic method may be equally or more effective but that further experimentation may be required for a more confident conclusion; three wavelengths do not uniquely determine a mixture's full spectrum, since metamers can share absorbance at the measured wavelengths while differing at others.