Cart-Pole Control on Cart-Pole
100Success RateGT Dense (Oracle)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GT Dense (Oracle)Reward Signal Type=Hand-crafted, incremental reward, RL Algorithm=SAC, Random Seeds=3, Evaluation Episodes=52026.07 | 100 | — | — | |
| GT Sparse (Oracle)Reward Signal Type=Binary success/failure signals, RL Algorithm=SAC, Random Seeds=3, Evaluation Episodes=52026.07 | 100 | — | — | |
| CLIPReward Signal Type=Cosine similarity between CLIP embeddings, RL Algorithm=SAC, Random Seeds=3, Evaluation Episodes=52026.07 | 100 | — | — | |
| RL-VLM-FReward Signal Type=VLM preference-based scalar reward, RL Algorithm=SAC, Random Seeds=3, Evaluation Episodes=52026.07 | 100 | — | — | |
| VLM-AR3LReward Signal Type=VLM-AR3L reward model, RL Algorithm=SAC, Random Seeds=3, Evaluation Episodes=52026.07 | 100 | — | — | |
| Direct BACservo rate (hz)=502015.12 | 18 | 20 | — | |
| Indirect BACservo rate (hz)=252015.12 | 15 | 686 | 47,500 | |
| Indirect BACservo rate (hz)=172015.12 | 14 | 760 | 35,500 | |
| Direct BACservo rate (hz)=172015.12 | 11 | 1,100 | 120,000 | |
| Indirect BACservo rate (hz)=502015.12 | 6 | 860 | 110,000 | |
| Direct BACservo rate (hz)=252015.12 | 3 | 2,300 | 380,000 |