RoboProcessBench - Phase Recognition: leaderboard
Metric: Accuracy (%) on 1,274 questions asking which coarse process phase a manipulation is in, from a single frame (4 options, chance 25.0); zero-shot multiple-choice VQA on held-out robot manipulation recordings from GM-100, RH20T, REASSEMBLE and AIST-Bimanual (ProcessData-Eval), temperature 0.01; higher is better. Source: arxiv.org. Saturation forecast: Around 2035. 14 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | InternVL3.5-8B | 37.4 |
| 2 | Claude Sonnet 4.6 | 31.9 |
| 3 | InternVL3-8B | 31.3 |
| 4 | GPT-5.4 Mini | 30.9 |
| 5 | GPT-4o | 29 |
| 6 | Qwen 3 VL 32B | 28.4 |
| 7 | Qwen 2.5 VL 7B Instruct | 26.6 |
| 8 | Claude Haiku 4.5 | 24.8 |
| 9 | InternVL3-38B | 24.6 |
| 10 | GLM-4.6V | 21.1 |
Interactive version: theaggregate.ai/benchmark?slug=roboprocessbench-phase-recognition · How It Works · Data refreshed daily, snapshot 2026-09-29.