EmbodimentSemantic - Eye-in-Hand: leaderboard
EmbodimentSemantic (2026) asks vision-language models to write the spatial scene graph of a simulated LIBERO-Spatial robot manipulation frame as directed object-relation-object triplets over seven objects and eight relations, checked against simulator-derived ground truth. This board uses the moving wrist camera, whose graphs cover only visible objects; the score is exact-triplet F1 averaged over the ten manipulation tasks, in percent, higher is better.
Metric: Mean per-task triplet F1 (%; directed object-relation-object triplets predicted from single LIBERO-Spatial eye-in-hand (wrist camera, graphs filtered to visible objects) frames against simulator-derived ground-truth scene graphs over a fixed vocabulary of 7 objects and 8 directed relations; exact triplet F1, invariant to swapping the two identical bowls, averaged over the 10 manipulation tasks; one fixed prompt, valid triplets only; higher is better). Source: arxiv.org. Saturation forecast: Around September 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 37.73 | #94 |
| 2 | Qwen 3 VL 8B Instruct | 21.08 | #981 |
| 3 | InternVL3-78B | 21.01 | #738 |
| 4 | gemma-4-E4B-it | 14.72 | #1469 |
| 5 | Molmo2-8B | 14.59 | #1483 |
| 6 | Nemotron Nano 12B v2 VL | 12.7 | #2231 |
| 7 | Qwen 2.5 VL 7B Instruct | 11.47 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=embodimentsemantic-eye-in-hand · How It Works · Data refreshed daily, snapshot 2026-11-13.