Failure Detection and Reasoning on RLBench
91.7Detection AccuracyARMOR
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ARMOR2026.02 | 91.7 | 71.8 | 80.2 | |
| SFT-S+D2026.02 | 72.6 | 55 | 64.6 | |
| SFT-D2026.02 | 64 | 46 | 60.6 | |
| Claude-3.7shots=32026.02 | 56.1 | 47.3 | 52.6 | |
| Claude-3.7shots=02026.02 | 42 | 37.2 | 33.6 | |
| Qwen2.5-VLshots=02026.02 | 37.6 | 25.5 | 35.3 | |
| Cosmos-Reasoning2026.02 | 31.7 | 22 | 14 | |
| LLaVA-NeXT2026.02 | 6.7 | 34.6 | 3.2 |