LabRobFail-Bench - Task Understanding: leaderboard
LabRobFail-Bench (Dalian University of Technology and collaborators, 2026) tests robot failure analysis in simulated chemical self-driving laboratories, where control, physics and semantic failures are injected into manipulation videos. This task asks the model to break the video into its sequence of atomic lab actions. Top-1 accuracy in percent on the seen-environment test split, higher is better.
Metric: Top-1 accuracy (%; decomposing the video into its atomic lab actions; robot-manipulation videos from simulated chemical self-driving laboratories with injected control, physics and semantic failures, seen-environment test split; answers from simulator ground truth; zero-shot VLM). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 45.16 | #123 |
| 2 | Gemini 2.5 Flash | 38.92 | #453 |
| 3 | Gemini 2.0 Flash | 26.98 | #704 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=labrobfail-bench-task-understanding · How It Works · Data refreshed daily, snapshot 2026-11-13.