Closed-loop feedback is a foundational requirement for self-driving laboratories (SDLs), yet large fraction of the instruments found in typical wet-lab environments remain resistant to automation. Micropipettes, hot plates, syringe pumps, centrifuges, and many benchtop characterization tools were designed for human operation: their operational state is communicated through digit wheels, dials, gauges, indicator lights, or liquid levels that a human eye reads effortlessly but that are not digitally accessible. Conventional solutions rely on dedicated feedback hardware, i.e., encoders, limit switches, photodetectors, load cells, that is effective but instrument-specific, often expensive, and rarely transferable across different equipment geometries or display types. This creates a persistent bottleneck for researchers who wish to automate existing laboratory equipment rather than replace it wholesale with proprietary, digitally-native alternatives.

This work introduces vision-language models (VLMs) as universal, software-defined closed-loop sensors for laboratory instrument automation. In the proposed architecture, an off-the-shelf USB camera observes visible operational state of an instrument, a pretrained VLM interprets that image using task-specific prompts, and the extracted state is passed to a host control layer that issues corrective actuation commands. Because the VLM does the interpretive work through prompting rather than through instrument-specific training, the same perceptual pipeline can in principle be redirected at any visually-readable instrument state, a digit wheel, a liquid meniscus, an indicator light, without retraining, custom electronics, or hardware-specific computer-vision development.

We demonstrate this concept using automated micropipette volume adjustment as a representative case study, since open-loop gear-driven actuation of a pipette's volume-setting mechanism is fundamentally unreliable: backlash, static friction, mechanical compliance, ratchet behaviour, and microstepping nonlinearity all introduce small but systematic positioning errors that motor commands alone cannot detect or correct. A camera was mounted on a 3D-printed fixture at a fixed standoff distance from the pipette's digit display; the VLM was prompted to read the displayed volume from each captured frame, and the host controller iteratively issued corrective motor commands until the observed and target volumes converged.

Open-loop actuation, evaluated across target settings from 20 to 100 µL, produced a mean absolute setting error of approximately 0.56 µL, with individual errors reaching up to 2.6 µL. The VLM-based closed-loop strategy, by contrast, converged to the target displayed volume within the resolution of the pipette display for every tested setting. Generalization was confirmed across three micropipette systems from different manufacturers (Four E's Scientific Precipette, IKA PETTE, and Socorex Acura), with same VLM prompt and controller architecture requiring no task-specific retraining, only the physical mounting fixture needed adaptation. Convergence was rapid and practical: across five evaluated pipette configurations, the median number of feedback iterations was 2–3, with an overall range of 1–5, variation attributable to differences in pipette geometry, display mechanics, and imaging conditions. The approach is also notably low-cost: the sensing hardware (camera plus 3D-printed fixture) totals approximately USD 20, and VLM inference costs approximately USD 0.000252 per call, keeping the total inference cost per volume-setting operation well below one cent.

These results establish that pretrained VLMs can convert human-readable, analog instrument states into machine-readable feedback without dedicated sensors, custom-electronics, or instrument-specific vision pipelines, offering a low-cost, generalizable route for retrofitting legacy and low-cost laboratory equipment into autonomous, closed-loop SDL workflows. Beyond pipette displays, the same architecture is directly extensible to liquid-level monitoring, vial detection, colour-change observation, meniscus tracking, and indicator-light reading, positioning VLM-based sensing as a practical, software-defined bridge between visually interpretable laboratory processes and fully autonomous experimentation.