WetRobo: A Reproducible Robot Kit for Coding Agents in Biological Laboratories
Yuna Oikawa, Kei Endo, Takanori Uzawa, Yunzhe Zhang, Manan Anjaria, Lerrel Pinto, Sherry Yang, Koji Tsuda
cs.AI · cs.RO
Quick verdict
WetRobo: a coding agent adapts a robot kit across two real wet labs and beats a fine-tuned VLA on cross-lab cap transfer; single-run, single-model evidence tempers the claim.
Summary
WetRobo packages an AgileX Piper arm, specified wet-lab equipment (EYELA incubator, D-PBS bottle, TPP Petri dish), a base code stack, an AGENTS.md skill file, and fifteen teleoperated demonstrations per task so that a coding agent — rather than a per-laboratory trained policy — adapts the robot on site. Experimentalists issue natural-language prompts; the agent (OpenAI Codex CLI, gpt-5.6-sol) observes the lab and writes/executes Python using OpenCV, NumPy, SAM 3, MuJoCo, and Record3D RGB-D, leaving an 'evolved program' on a git branch. Across two laboratories differing in gripper, camera count, luminance (130.10 vs 106.84), and bench color, the agent lifted a Petri dish lid, removed a bottle cap in both labs (23.0M–41.4M tokens; 48 min – 1 h 02 min to first grasp), and opened the incubator door in Lab X, repeatable in 3 min 27 s; the Lab Y door attempt failed on unreliable head-camera alignment. Against π0.5 fine-tuned on 25 Lab X demonstrations, the agent scored 10/10 in both labs where the VLA fell to 0/10 in Lab Y. Main limitations: one agent/model, one adaptation run per lab-task, success rates reflecting same-setting re-execution rather than adaptation reliability, and an asymmetric VLA comparison (zero-shot transfer vs online adaptation). Code and evolved branches are on GitHub.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 5
● low 0
4 strengths · 4 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The WetRobo kit lets biological experimentalists delegate manipulation tasks to a coding agent without collecting local teleoperation data or training neural networks.
- Codex (gpt-5.6-sol) autonomously evolved programs that lifted a Petri dish lid, removed a bottle cap in two laboratories, and opened an incubator door in real-world labs.
- On bottle-cap lifting, the coding agent succeeded 10/10 in both Lab X and Lab Y, whereas π0.5 fine-tuned on 25 Lab X demonstrations achieved 10/10 in Lab X but 0/10 in Lab Y.
- Evolved programs and recorded actions feed back into AGENTS.md and shared branch exemplars, making each installation reusable procedural knowledge expected to cut adaptation cost in new labs.
Strengths
Cross-laboratory evaluation with documented confounds
Table I quantifies lab differences (luminance 130.10 vs 106.84, NYU string-driven vs Piper-native gripper, 2 vs 3 cameras) and Table II contrasts how the evolved programs diverged (torque-stop vs MuJoCo collision pre-checks; effort+opening vs visual+opening grasp evidence), rather than hiding environmental variation.
Transparent reporting of failure
The Lab Y incubator-door failure is reported with its mechanism (unreliable head-camera alignment triggering a pre-contact stop and return home), and Fig. 8 maps pre-close orientation spread across ten attempts against tool additions — an unusually candid account for a systems paper.
Releasable artifacts of code evolution
Mechanistic depth beyond success rates
Weaknesses
Adaptation reliability is unmeasured
Asymmetric VLA baseline
Abstract phrasing obscures partial door-task success
Proprietary-model dependence and generatively edited results figure
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WetRobo?
WetRobo is a robot kit for biological laboratories that combines a low-cost Piper robot arm, specified lab equipment (an incubator, a reagent bottle, a Petri dish), a base software stack, an AGENTS.md instruction file, and 15 teleoperated demonstration videos per task. Instead of training a robot policy per lab, a coding agent such as OpenAI Codex observes the local lab, writes and executes robot-control programs, and adapts the kit on site from a natural-language task description.
Did the coding agent really beat the VLA (π0.5)?
On the bottle-cap task, yes in this specific test: the coding agent succeeded in 10 of 10 runs in both labs, while π0.5 fine-tuned on 25 demonstrations from Lab X scored 10/10 in Lab X but 0/10 in the new lab. Caveats: it is one task and one VLA, the VLA was tested without any adaptation in the new lab while the agent could observe and rewrite code on site, and each agent adaptation was run only once.
Which tasks succeeded and which failed?
Petri-dish lid lifting succeeded; bottle-cap removal succeeded in both laboratories (Lab X and Lab Y); incubator-door opening succeeded in Lab X — repeatable in 3 min 27 s after a 2 h 40 min first run — but the single Lab Y attempt failed because head-camera alignment was unreliable, leaving the door closed.
Is the WetRobo code and data available, and what did it cost to run?
Yes — the repository, demonstrations, and the evolved programs (as git branches per task/lab) are at github.com/tsudalab/WetRobo. The cap trials consumed about 23 million tokens in Lab Y and 41 million in Lab X, taking roughly 49 minutes and 1 hour respectively to the first grasp; results depend on the proprietary gpt-5.6-sol model, which may limit exact reproduction.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures could not be visually inspected (text extraction provides captions only), so Fig. 6 token/gripper traces, Fig. 8 orientation spread, and Fig. 9 regression quality were judged from captions and prose.
- The provided abstract is truncated mid-sentence ('...let a coding agent'), and some fine-grained numbers (e.g., Fig. 8 attempt-level orientations) depend on figure content not present in the text.
- The accompanying video referenced in Sections III-A/III-B was not reviewed; claims about execution quality rest on the written description.
- References dated 2026 (e.g., SAM 3, CaP-X, ENPIRE, LIBERO-Plus) could not be verified against the network per review constraints.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.