Interaction-aware object filtering on IA-Bench 1.0 (held-out test)
80.2AccuracySPARC (Final)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SPARC (Final)Category=Ours (Final), Scoring rule=Ours (Final)2026.06 | 80.2 | 77.6 | 5.6 | 3.5 | |
| SPARC (+ Phase aware movement)Category=Ablation, Scoring rule=+ Robot filter + Phase aware movement2026.06 | 77.8 | 73.5 | 6.4 | 3.7 | |
| SPARC (+ Robot filter)Category=Ablation, Scoring rule=+ Mean movement + Robot filter2026.06 | 72.7 | 54.9 | 10.1 | 5.9 | |
| Gemini Robotics ER 1.6Category=External, Scoring rule=Gemini Robotics ER 1.62026.06 | 71.7 | — | — | — | |
| SPARC (+ Mean movement)Category=Ablation, Scoring rule=Det. confidence (LLMDet) + Mean movement2026.06 | 69.7 | 51.4 | 11.7 | 6.6 | |
| Det. confidence (OWLv2)Category=Baselines, Scoring rule=Det. confidence (OWLv2)2026.06 | 62.2 | 0.2 | 29.7 | 0.2 | |
| Det. confidence (LLMDet)Category=Ablation, Scoring rule=Det. confidence (LLMDet)2026.06 | 58.1 | 26.6 | 21.9 | 11.5 | |
| Motion magnitudeCategory=Baselines, Scoring rule=Motion magnitude2026.06 | 46.3 | 0 | 65.7 | 47.6 | |
| FSD – Traj. lengthCategory=Baselines, Scoring rule=FSD – Traj. length2026.06 | 38.6 | 33.1 | 27.9 | 3.2 | |
| Qwen3.6-30B-A3BCategory=External, Scoring rule=Qwen3.6-30B-A3B2026.06 | 6 | — | — | — |