Multi-task robot manipulation on LIBERO-PRO-Long unseen benchmark
42Language ErrorOpenVLA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| OpenVLABackbone=OpenVLA, Evaluation Protocol=zero-shot, Sampling Strategy=greedy decoding2026.02 | 42 | 26.6 | 3.2 | 0 | 18 | |
| OpenVLA + Top-pBackbone=OpenVLA, Evaluation Protocol=zero-shot, Sampling Strategy=Top-p (p=0.9)2026.02 | 44.2 | 28.2 | 3.6 | 0 | 19 | |
| OpenVLA + Top-kBackbone=OpenVLA, Evaluation Protocol=zero-shot, Sampling Strategy=Top-k (k=40, t=0.7)2026.02 | 44.6 | 26 | 4 | 0 | 18.7 | |
| OpenVLA + SamplingBackbone=OpenVLA, Evaluation Protocol=zero-shot, Sampling Strategy=Sampling (t=1.0)2026.02 | 44.8 | 27.6 | 3.2 | 0 | 18.9 | |
| SCALEBackbone=OpenVLA, Evaluation Protocol=zero-shot2026.02 | 51.2 | 30 | 4.8 | 0 | 21.5 | |
| π0-FAST + SamplingBackbone=π0-FAST, Evaluation Protocol=zero-shot, Sampling Strategy=Sampling (t=1.0)2026.02 | 74.8 | 46.2 | 13.8 | 2.4 | 34.3 | |
| π0-FAST + Top-pBackbone=π0-FAST, Evaluation Protocol=zero-shot, Sampling Strategy=Top-p (p=0.9)2026.02 | 75.4 | 44.8 | 14.2 | 2.6 | 34.3 | |
| π0-FAST + Top-kBackbone=π0-FAST, Evaluation Protocol=zero-shot, Sampling Strategy=Top-k (k=40, t=0.7)2026.02 | 75.6 | 44.6 | 14.6 | 2.6 | 34.4 | |
| π0-FASTBackbone=π0-FAST, Evaluation Protocol=zero-shot, Sampling Strategy=greedy decoding2026.02 | 78 | 49 | 13.6 | 2.2 | 35.7 | |
| SCALEBackbone=π0-FAST, Evaluation Protocol=zero-shot2026.02 | 84 | 51.8 | 15.8 | 3.4 | 38.8 |