Pairwise Preference Prediction on RBM-EVAL-OOD Different Quality Trajectory Pairs
100Preference Accuracy (USC Franka)ROBOMETER
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ROBOMETERTrajectory pair type=Different task instructions, Number of comparisons=500 per dataset2026.03 | 100 | 89.8 | 99 | 98.2 | 98.4 | 100 | 97.6 | |
| ROBOMETER2026.03 | 75 | 79.4 | 76.2 | 88.9 | 85.4 | 90 | 82.5 | |
| RL-VLM-FTrajectory pair type=Different task instructions, Number of comparisons=500 per dataset2026.03 | 70.7 | 54.7 | 64 | 73.3 | 55.3 | 72.7 | 65.1 | |
| RL-VLM-Fmodel=GPT-4o-mini2026.03 | 52.1 | 54.4 | 66.7 | 48.6 | 54.4 | 56.7 | 55.5 |