WetRobo is a robot kit that can be transferred between laboratories without requiring teleoperation or neural‑network training. The kit comprises a single robot arm, laboratory hardware (an incubator, a capped reagent bottle, and a Petri dish), the existing arm‑control code, recorded teleoperation demonstrations for each task, and a generic AGENTS.md skill file. A biologist simply provides a natural‑language instruction; a coding agent observes the local lab, writes and runs a program, and calls external tools as needed for adaptation.
Using OpenAI Codex (gpt-5.6-sol), we demonstrated three real‑world tasks: lifting a Petri dish lid, removing a bottle cap, and opening the incubator door. The coding agent succeeded on the cap task in both Lab X and Lab Y, whereas a Vision‑Language‑Action (VLA) policy fine‑tuned on Lab X demonstrations only worked there and failed to transfer to Lab Y.
These results suggest a practical path for laboratory robotics: instead of training a separate policy for each lab, distribute a reproducible kit and let a coding agent adapt it on‑site. The code, demonstrations, and evolved programs are available at https://github.com/tsudalab/WetRobo.
Review