A new robot kit lets an AI coding agent handle basic lab tasks without any lab-specific training data.
The system, called WetRobo, pairs a robot arm with real lab gear, an incubator, a reagent bottle with a cap, and a Petri dish, plus recorded teleoperation demos and a skill file the coding agent can read. A researcher describes a task in plain language, and the coding agent looks at the lab, then writes and runs the code needed to carry it out, pulling in outside tools if the setup differs from what it has seen before. The team tested this with OpenAI Codex, listed in the paper as gpt-5.6-sol, on three physical tasks: lifting a Petri dish lid, removing a bottle cap, and opening an incubator door, across two separate labs. The coding agent completed the cap task in both labs, while a vision-language-action policy trained only on the first lab's setup worked there but failed in the second.
That gap is the real finding. Most lab robotics work trains one policy per environment, which means retraining every time a bench, incubator model, or bottle shape changes. Treating the robot's actions as code that an agent writes and rewrites on the spot, the same way coding agents already adapt to new codebases, sidesteps that retraining cycle entirely.
Three tasks, lifting a lid, twisting a cap, opening a door, is a long way from running an actual experiment. Pipetting, sample prep, and recovering when something goes wrong will be a much harder test of whether this approach scales past a demo.