RoboQuest Benchmark Shows Frontier Models Fail Most Kitchen Inspection Tasks

What changed

RoboQuest, open sourced 5 October 2026 on GitHub, is a benchmark where a mobile manipulator must search, physically inspect, and decide when kitchen style tasks are complete. The README reports evaluation of five frontier multimodal models through a shared visuomotor interface, with the best succeeding in roughly 23.2% of episodes across ten RoboCasa365 kitchen tasks. The project ships charts such as success-by-task.png and cites arXiv 2610.10388.

Why it matters

Food safety, clinical lab prep, and biomanufacturing all require agents that gather evidence through contact and micro inspection, not only VQA style guesses. A public benchmark with low success rates is a useful antidote to marketing slides claiming general purpose physical agents are production ready.

Who is affected

Biotech automation engineers, lab robotics vendors, and QA leaders in regulated wet labs. Model providers should expect procurement teams to ask for RoboQuest style scores before approving autonomous inspection in GMP adjacent settings.

What to do next

If you deploy mobile manipulators near biological samples, run at least one RoboQuest task suite in simulation before expanding autonomy. Treat sub 30% frontier success as a default risk flag, not an exception.

What to watch

Whether labs publish independent runs on the released 5,000 demonstration trajectories and whether regulators reference physical inspection benchmarks in 2026 agent guidance.

Image: Declare Lab

Sources

• Primary — Declare Lab, RoboQuest repository (5 October 2026). Task design, model lineup, reported success rates.
• Secondary — arXiv, RoboQuest preprint 2610.10388 (2026). Academic framing and citation details.