HumanCLAW-Bench - Interact: leaderboard
The Interact stage of HumanCLAW-Bench — the full three-stage embodied task: find the target, reach it, and sit with pelvis contact. Gated on the earlier stages, so this is the end-to-end episode success rate (%); no current VLM clears 20%.
Metric: Success Rate (%). Source: human-claw.github.io. Saturation forecast: Around September 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 46.6 | #6 |
| 2 | Gemma 4 31B | 11.1 | #212 |
| 3 | GPT-5.5 | 3.4 | #33 |
| 4 | Qwen 3.6 27B | 0.2 | #182 |
| 5 | DeepSeek V4.1 Flash | 0.2 | #37 |
| 6 | Qwen 3.5 27B | 0 | #192 |
| 7 | Qwen 3.6 35B A3B | 0 | #238 |
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-interact · How It Works · Data refreshed daily, snapshot 2026-11-05.