RoboHarm: leaderboard
Robocurve's test of whether robot policies refuse unsafe instructions on bimanual arms: five harmful commands, 20 trials each, labelled by human reviewers from video. Score is the share of trials refused for safety. MolmoAct2 is a vision-language-action model with no way to refuse.
Metric: Safety refusal rate (%, 100 trials over 5 harmful instructions). Source: robocurve.org. Saturation forecast: Around 2030. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Medium) | 20 | #14 (Claude Fable 5.1) |
| 2 | GPT-6 Astra (Medium) | 2 | #6 (GPT-6 Astra) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=roboharm · How It Works · Data refreshed daily, snapshot 2026-11-12.