Embodied3DBench - Grounding (2D): leaderboard

Embodied3DBench (2026) Grounding (2D) task on simulated robotic tabletop scenes: object and part pointing and bounding boxes in the image plane, predicted zero-shot by a vision-language model. Normalized 0-100 score derived from box IoU or point distance; higher is better.

Metric: Grounding (2D) task (object and part pointing and bounding boxes in the image plane), normalized score (0-100): IoU of predicted boxes or the L2 error of predicted points (pixels in 2D, metres in 3D) mapped to 0-100, on the simulated robotic tabletop scenes of Embodied3DBench (21K QA pairs over 4.8K object instances), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 2.5 VL 72B Instruct62.8#841
2GPT-553.7#178
3Qwen 3 VL 4B Instruct49.6#1332
4Qwen 2.5 VL 7B Instruct48.5#1902
5Claude Sonnet 436.8#410
6GPT-4o28.7#751
7Gemini 2.5 Flash21.8#463
8Gemini 2.5 Pro20.5#276
9InternVL3.5-8B19.9#1604
10InternVL3-38B17.8#939
11Doubao-Seed-1.6Thinking17.6
12Doubao-1.5-thinking-vision-pro16.9#344
13Seed-1.6Non-reasoning15.8#469

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=embodied3dbench-grounding-2d · How It Works · Data refreshed daily, snapshot 2026-11-21.