RoboGraphBench - Tabletop: leaderboard
RoboGraphBench (Wang and Li, 2026) compiles long-horizon robot tasks into symbolic scene graphs whose task-relevant state must be tracked through exploration, action failures and interventions, and runs models as closed-loop high-level planners. This board covers 399 tabletop episodes from 57 scenes, each in a baseline and six intervention conditions. Task success rate in percent; higher is better.
Metric: Task success rate (%; episodes that reach the goal and stop correctly, averaged over the baseline and six intervention conditions; semantic closed-loop evaluation on partially observed symbolic scene graphs, at most 100 steps; 399 tabletop episodes from 57 scenes). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 78.4 | #48 |
| 2 | Gemini 3.1 Pro (Preview) | 62.7 | #91 |
| 3 | Claude Opus 4.7 | 56.6 | #75 |
| 4 | DeepSeek V4 Pro | 49.1 | #181 |
| 5 | GLM-5.2 | 43.6 | #137 |
| 6 | Qwen 3.7 Plus | 24.3 | #203 |
| 7 | GPT-5.4 | 23.1 | #121 |
| 8 | Qwen 3.6 Plus | 21.1 | #217 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=robographbench-tabletop · How It Works · Data refreshed daily, snapshot 2026-11-12.