HumanCLAW-Bench - Find: leaderboard
The Find stage of HumanCLAW-Bench — a frozen off-the-shelf VLM drives a simulated humanoid over 1,218 episodes and must get the target object rendered in its own ego view and acknowledge it; FindSR success rate (%).
Metric: Success Rate (%). Source: human-claw.github.io. Saturation forecast: Around April 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 75.5 | #6 |
| 2 | DeepSeek V4.1 Flash | 68.1 | #37 |
| 3 | Gemma 4 31B | 58.1 | #212 |
| 4 | GPT-5.5 | 55.1 | #33 |
| 5 | Qwen 3.6 27B | 51 | #182 |
| 6 | Qwen 3.6 35B A3B | 44.6 | #238 |
| 7 | Qwen 3.5 27B | 37.8 | #192 |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=humanclaw-bench-find · How It Works · Data refreshed daily, snapshot 2026-11-05.