RoboSpatial-Home
Multimodal
RoboSpatial-Home is the 350-question indoor evaluation split of RoboSpatial (123 configuration, 105 compatibility and 122 context questions), where configuration and compatibility are yes/no spatial questions and context questions require pointing to a free-space location judged by point-in-mask. This definition is item accuracy pooled over all 350 questions; the RoboSpatial-Eval leaderboard's overall total is instead the mean of the three category scores.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-4B | 67.4 |
| Gemma 4 12B IT | 66.9 |
| GPT-5.6 Luna | 63.7 |
| LFM3-VL-2B | 63.4 |
| LFM3-Omni-2B | 49.1 |
| Qwen3.5-2B | 47.7 |
| Gemma 4 E2B IT | 46.9 |
| Gemini 3.5 Flash-Lite | 41.4 |
| Phi-4-multimodal-instruct | 41.1 |