Robo-bench — Humanoid
Agents · 2026-10-13
Mean ordered-milestone progress on Robo-bench's 11 humanoid locomotion and manipulation tasks: a Unitree G1 with Dex3 hands, a head camera and no balance or walking controller, so the agent commands every motor itself. One three-minute real-time episode per task.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 1.8 |
| GPT-6.1 Sol | 1.8 |
| Muse Wonder Full | 1.8 |
| GPT-6 Astra | 0.9 |
| Claude Fable 5.1 | 0.0 |
| Gemini 4 Argon | 0.0 |
| Grok 4.7 | 0.0 |