Embodied3DBench - 2D Average: leaderboard

Embodied3DBench (2026) low-level embodied spatial intelligence in the image plane: a vision-language model grounds parts and objects, matches points across views, and predicts affordances, manipulation trajectories and grasp points on simulated robotic tabletop scenes. Mean of five normalized 0-100 task scores derived from box IoU and point distance; higher is better.

Metric: Mean of the five 2D task scores, each a 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 3 VL 4B Instruct56.5#1332
2Qwen 2.5 VL 72B Instruct55.5#841
3GPT-554.4#178
4Qwen 2.5 VL 7B Instruct43#1902
5Claude Sonnet 441.2#410
6GPT-4o36.6#751
7Gemini 2.5 Flash31.8#463
8Gemini 2.5 Pro31.4#276
9InternVL3.5-8B28.9#1604
10InternVL3-38B27.4#939
11Doubao-1.5-thinking-vision-pro27.4#344
12Doubao-Seed-1.6Thinking27.2
13Seed-1.6Non-reasoning25.8#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-2d-average · How It Works · Data refreshed daily, snapshot 2026-11-21.