# A Frugal Self-Driving Lab on Opentrons OT-2: Closed-Loop Color Matching and Acid-Base "Battleship"

## Jon Potter1 , Yixuan Yin1, Zhishan Liu1, Shichen Liu1, John Kitchin1, Stefan Bernhard1, Joshua D. Kangas1

##  (1 Carnegie Mellon University)

**Purpose** - We introduce a low-cost, modular self-driving lab (SDL) built on an Opentrons OT-2 that teaches and prototypes closed-loop experimentation. The system supports two applications: (1) dye color matching and (2) an acid-base "Battleship" game with algorithmic players. This work has not been the subject of a podium presentation and includes new results from a 2025 deployment for \~100 high school students.

**Experimental procedures** -   
Hardware: OT-2 liquid handler; 96-well plates; six fixed downward-facing \~$70 cameras on 3D-printed mounts; white index-card background. 

Software: a persistent OT-2 protocol watches a local ".jsonx" file every 5s; a server writes action lists and reads logs, avoiding repeated 6-min reboots typical of per-run uploads on Opentrons systems. 

Color sensing: one-time manual well-center calibration; radial Gaussian pixel sampling per well; online clustering with a 10-RGB-unit threshold; the centroid of the largest cluster is reported; plate-wide lighting is normalized by empty-well baselining. 

Modeling/active learning: We fit three Gaussian-process models (one per RGB channel) using a Constant x RBF plus White kernel. Recipe volumes obey a simplex constraint (total 200 μL). Next experiments are chosen by minimizing color-distance with an uncertainty bonus, optimized with multi-start trust-constr. 

Activities: (i) color matching both with three food dyes and acid/base/pH indicator as well as an 11-guess limit per row; (ii) Battleship with water "ocean," acid "ships," and pH-indicator "missiles," with hits/misses classified by camera; teams extend a Python base class ("ShotSelection") to build strategies;

**Summary of data** -   
Color matching: Across >=500 wells, the camera pipeline produced consistent RGB reads despite glare, refraction, and meniscus artifacts. In round 1 of dye matching, the majority of 12 student groups outperformed the AI; parity emerged in round 2; in rounds 3-4 the GP-driven AI required fewer guesses to reach a <=20-RGB-unit match threshold, while its poor round-1 start kept its overall mean higher than the students', the median favored the AI. 

Battleship: The Battleship tournament delivered fully automated scoring and a complete single-elimination bracket; student bots improved markedly over the baseline heuristic.

**Conclusion** - A frugal SDL can deliver reliable closed-loop experimentation on an OT-2 without expensive optics or vendor-locked workflows, and it generalizes across assay-like tasks such as quantitative color and  titration-style detection while also proving to be powerful tangible mechanism for teaching active learning models. Open-source code and CAD are provided.

**Next steps and future experiments** - Investigate stronger resilience to ambient light changes; evaluate alternate chemical color gradients; compare acquisition functions and optimizers.