Robotic Manipulation on LIBERO closed-loop evaluation
99Spatial SuccessSnapFlow
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| SnapFlowParams=3.0B, Steps=12026.04 | 99 | 100 | 99 | 97 | 98.75 | 0.0077 | 0.9916 | 83 | 3.3 | |
| Baseline (Euler)Params=3.0B, Steps=102026.04 | 98 | 100 | 96 | 97 | 97.75 | 0.0117 | 0.9885 | 274 | 1 | |
| π0Params=3.0B, Steps=102026.04 | 97.4 | 98.4 | 97.6 | 93 | 96.6 | — | — | — | — | |
| Naïve 1-stepParams=3.0B, Steps=12026.04 | 96 | 99 | 98 | 94 | 96.75 | 0.0089 | 0.9911 | 81 | 3.4 | |
| OpenVLAParams=7.0B, Steps=AR2026.04 | 84.9 | 88.4 | 79.2 | 53.7 | 76.55 | — | — | — | — | |
| VLA-AD (with-Qwen)Model Role=Student, Number of parameters=158M, Teacher=OpenVLA-7B, VLM supervision=with-Qwen2026.05 | 79.5 | 62.5 | 79.5 | — | 73.8 | — | — | — | — | |
| OpenVLA-7BModel Role=Teacher, Number of parameters=7B2026.05 | 79 | 62 | 81 | — | 74 | — | — | — | — | |
| Octo-BaseParams=93M, Steps=102026.04 | 78.9 | 85.7 | 84.6 | 51.1 | 75.08 | — | — | — | — | |
| Diff. PolicyParams=—, Steps=1002026.04 | 78.3 | 92.5 | 68.3 | 50.5 | 72.4 | — | — | — | — | |
| VLA-AD (no-Qwen)Model Role=Student, Number of parameters=158M, Teacher=OpenVLA-7B, VLM supervision=no-Qwen2026.05 | 77 | 62 | 61 | — | 66.7 | — | — | — | — |