Behavior Cloning on DeepMind Control Suite (seen/unseen embodiments)
49.1Hopper Hop ScoreMeta-Controller
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Meta-Controllern-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 49.1 | 87.2 | 82.5 | 91.7 | 67.3 | 56.1 | 50.8 | 84.3 | 71.1 | |
| MetaMorphn-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 33.4 | 81.8 | 54.1 | 73 | 29.7 | -4.1 | 3.8 | 31.8 | 37.9 | |
| MTGv2n-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 21.6 | 83.1 | 40.7 | 68.2 | 29.1 | -1.4 | 4.6 | 35.5 | 35.2 | |
| L2Mn-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 4.5 | 45.1 | 4 | 50.6 | 65.9 | 3.7 | 15.5 | 40 | 28.7 | |
| MT-DTn-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 3.2 | 64.3 | 10.1 | 53.4 | 46.2 | 10.2 | 8.4 | 86.4 | 35.3 | |
| PDT+PEFTn-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 2.3 | 61.7 | 12.3 | 54.9 | 52.7 | -1.7 | 9.1 | 74.1 | 33.2 | |
| FS-DTn-shot demonstrations=5, adaptation protocol=from-scratch, backbone architecture=6-layer transformer2024.12 | 1.6 | 56.9 | 9.8 | 44.7 | 35.8 | -0.7 | 5.5 | 20.6 | 21.8 | |
| PDTn-shot demonstrations=5, adaptation protocol=prompting, backbone architecture=6-layer transformer2024.12 | 0.9 | 29.7 | 6.5 | 47 | 35.7 | 3.3 | 5.8 | 5 | 16.7 | |
| HDTn-shot demonstrations=5, adaptation protocol=fine-tuning, backbone architecture=6-layer transformer, fine-tuning iterations=10,0002024.12 | 0.8 | 51.6 | 13.7 | 56.4 | 38.7 | 1.2 | 8.5 | 3.3 | 21.8 |