Failure Detection and Reasoning on Maniskill
78.8Detection AccuracySFT-D
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SFT-D2026.02 | 78.8 | 0.644 | 74.3 | |
| Claude-3.7shots=32026.02 | 62.5 | 0.398 | 21.2 | |
| Qwen2.5-VLshots=02026.02 | 54.8 | 0.268 | 11.2 | |
| Claude-3.7shots=02026.02 | 53.8 | 0.36 | 13.3 | |
| Cosmos-Reasoning2026.02 | 44.2 | 0.359 | 20.7 | |
| LLaVA-NeXT2026.02 | 1 | 0.286 | 4.2 |