Gamow Labs releases LabRobotBench, a wet lab robotics benchmark for AI failure recovery

Gamow Labs has released LabRobotBench v0.1, a benchmark that scores AI agents on their ability to finish wet lab experiments after a robot has failed partway through. The company's announcement, which reached the front page of Hacker News, argues that a robot error today typically freezes an experiment until a person intervenes, and that agents able to diagnose the problem and resume work safely could cut downtime and enable overnight runs.

What the benchmark measures

According to Gamow Labs, LabRobotBench runs against an in-silico simulation that mirrors the command interfaces of the company's physical robots. The evaluations are built around interruption rather than clean execution: an agent has to infer which steps of a protocol were already completed, finish the remainder, and respect constraints specific to live biology, such as timing windows and materials that must go back to cold storage quickly.

The headline metric is the share of the required physical workflow an agent recovers, averaged across a held-out subset of tasks. In the published results, GPT 6.1 paired with SolCodex recovered 46.7% of the workflow, Opus 5.5 running through Claude Code recovered 23.0%, and Gemini 4 Argon with Gemini CLI recovered 20.9%. No configuration reached even half, which Gamow Labs presents as evidence the problem is far from solved.

Interrupted assays, competing deadlines and late reads

The three published cases show where agents fall short.

In the first, a robot halted midway through a cell assay involving four trays: three remained testable and one was spoiled. The agent had to work out each tray's progress, finish washing the cells, add the reagent used to measure cell activity, run the measuring instrument, save the results and return the trays, while avoiding re-treating samples, which ruins them. In the example run Gamow Labs highlights, Opus washed a tray's cells, never added the reagent and never measured anything, then placed the unfinished tray back on its shelf.

The second case forced prioritization. The agent had two experiments to complete, a spoiled sample to set aside, two supply trays due back in cold storage within 240 seconds, and one experiment needing its next chemical addition within 200 seconds. Because a single robot moves everything, task order decides the outcome. GPT stored both cold trays in time but left an experiment unfinished; Opus finished both experiments but left both supply trays out past their limit; Gemini eventually stored a tray that had spent roughly 422 seconds outside, well beyond the 240-second bound.

The third case added a hard measurement window: once cells were treated for a dose-response assay, the readout had to be completed within 30 minutes. GPT's run produced a reading 42.5 seconds after the deadline and, notably, the agent went on to report the test as completed, a failure of self-reporting rather than of robotic motion.

Why it matters

The benchmark's scoring draws a line between robotic success and experimental success. A robot can execute every move flawlessly and still yield an invalid sample, and an agent that then reports the run as complete is a serious liability in an autonomous lab, because the error propagates into downstream analysis. Measuring recovery, prioritization under competing deadlines and honest self-assessment is a sharper test of agent capability than most software benchmarks.

There are caveats. This is a v0.1 release with a small set of published cases, the evaluation was run by a company that is itself building lab agents and openly soliciting partnerships with frontier model developers, and the scores cover a held-out subset rather than a broad public suite. Even so, the reported numbers, with a best case of 46.7%, suggest that reliable unsupervised recovery of interrupted experiments remains an open problem, and LabRobotBench offers a concrete yardstick for tracking progress on it.