Robot Manipulation on LIBERO Spatial Object Goal Long
97.5Spatial Success ScoreOriginal (Un-compressed)
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Original (Un-compressed)VLA model=π0.5, Compression rate=None (original), VLM backbone=PaliGemma2026.06 | 97.5 | — | — | — | — | 100 | 97 | 96.5 | 97.75 | 96.47 | |
| Q-VGMCritic Usage Type=Training-time fine-tuning, Method Category=Training-time critic fine-tuning2026.06 | 96 | — | — | — | — | 95 | 95 | 84 | 92.5 | — | |
| EinSortVLA model=π0.5, Compression rate=0.4, VLM backbone=PaliGemma, Activation calibration episodes=642026.06 | 96 | — | — | — | — | 99.5 | 97.5 | 95.5 | 97.13 | 95.72 | |
| SemiVLA + Selective LoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Semi-supervised adaptation, Fine-tuning method=Selective LoRA2026.06 | 92.4 | — | — | — | — | 89.5 | 91.8 | 82.3 | 89 | — | |
| SemiVLA + LoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Semi-supervised adaptation, Fine-tuning method=LoRA2026.06 | 91.7 | — | — | — | — | 88.6 | 90.9 | 80.5 | 87.9 | — | |
| Test-time Q GuidanceCritic Usage Type=Test-time, Method Category=Test-time critic use2026.06 | 91 | — | — | — | — | 93 | 88 | 68 | 85 | — | |
| Q-Improved Action DistillationCritic Usage Type=Training-time fine-tuning, Method Category=Training-time critic fine-tuning2026.06 | 91 | — | — | — | — | 92 | 88 | 74 | 86.3 | — | |
| SemiVLA + QLoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Semi-supervised adaptation, Fine-tuning method=QLoRA2026.06 | 90.8 | — | — | — | — | 87.4 | 89.7 | 78.9 | 86.7 | — | |
| SemiVLA + AdapterSup.=10%, Backbone=OpenVLA, Adaptation Strategy=Semi-supervised adaptation, Fine-tuning method=Adapter2026.06 | 88.9 | — | — | — | — | 84.7 | 87.2 | 74.6 | 83.9 | — | |
| Test-time Q SelectionCritic Usage Type=Test-time, Method Category=Test-time critic use2026.06 | 88 | — | — | — | — | 91 | 85 | 64 | 82 | — | |
| OpenVLA + Selective LoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Supervised adaptation, Fine-tuning method=Selective LoRA2026.06 | 87.6 | — | — | — | — | 83.4 | 85.6 | 71.8 | 82.1 | — | |
| OpenVLA + LoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Supervised adaptation, Fine-tuning method=LoRA2026.06 | 86.8 | — | — | — | — | 82.1 | 84.9 | 70.3 | 81 | — | |
| OpenVLA + QLoRASup.=10%, Backbone=OpenVLA, Adaptation Strategy=Supervised adaptation, Fine-tuning method=QLoRA2026.06 | 85.9 | — | — | — | — | 80.5 | 83.7 | 68.9 | 79.8 | — | |
| VLA-PrunerFLOPs=39.80%, Speedup=1.633×2025.11 | 85.4 | — | — | — | — | 82.5 | 78.4 | 51.8 | 74.5 | — | |
| OpenVLA + AdapterSup.=10%, Backbone=OpenVLA, Adaptation Strategy=Supervised adaptation, Fine-tuning method=Adapter2026.06 | 84.6 | — | — | — | — | 78.2 | 81.5 | 65.7 | 77.5 | — | |
| EfficientVLA (T-only)FLOPs=58.12%, Speedup=1.334×2025.11 | 84.3 | — | — | — | — | 81.6 | 76.5 | 49.8 | 73.1 | — | |
| EfficientVLAFLOPs=41.30%, Speedup=1.491×2025.11 | 83.2 | — | — | — | — | 80.3 | 75.7 | 49.1 | 72 | — | |
| FitPruneFLOPs=47.71%, Speedup=1.397×2025.11 | 82.4 | — | — | — | — | 80.1 | 74.8 | 50.3 | 71.9 | — | |
| Initial policy (π0.5 few-shot SFT)Category=Initialization2026.06 | 82 | — | — | — | — | 84 | 78 | 56 | 75 | — | |
| OpenVLA Zero-shotSup.=0, Backbone=OpenVLA, Adaptation Strategy=Zero-shot2026.06 | 80 | — | — | — | — | 69.6 | 74 | 55.5 | 69.8 | — | |
| VTWFLOPs=55.17%, Speedup=1.325×2025.11 | 74.7 | — | — | — | — | 76.2 | 72.3 | 48.9 | 68 | — | |
| Diffusion-QLCritic Usage Type=Training-time fine-tuning, Method Category=Training-time critic fine-tuning2026.06 | 73 | — | — | — | — | 78 | 74 | 53 | 69.5 | — | |
| LatentLLMVLA model=π0.5, Compression rate=0.4, VLM backbone=PaliGemma, Activation calibration episodes=642026.06 | 66.5 | — | — | — | — | 88.5 | 37 | 37.5 | 57.38 | 53.92 | |
| SVD-LLMVLA model=π0.5, Compression rate=0.4, VLM backbone=PaliGemma, Activation calibration episodes=642026.06 | 60 | — | — | — | — | 79 | 32 | 27 | 49.5 | 45.05 | |
| Plain SVDVLA model=π0.5, Compression rate=0.4, VLM backbone=PaliGemma, Activation calibration episodes=None2026.06 | 27 | — | — | — | — | 34.5 | 8.5 | 1 | 17.75 | 15.26 | |
| ASVDVLA model=π0.5, Compression rate=0.4, VLM backbone=PaliGemma, Activation calibration episodes=642026.06 | 4 | — | — | — | — | 5.5 | 7.5 | 0 | 4.25 | 3.06 | |
| Pi0XPU Type=4090, Memory(GB)=24, Bandwidth(GB/s)=10002026.04 | — | 86 | 3,500 | 2.398 | 102.3 | — | — | — | — | — | |
| Pi0XPU Type=Thor, Memory(GB)=128, Bandwidth(GB/s)=2732026.04 | — | 86 | 3,400 | 1.282 | 246 | — | — | — | — | — | |
| Pi0XPU Type=Orin, Memory(GB)=64, Bandwidth(GB/s)=2042026.04 | — | 86 | 1,999 | 1.866 | 920.6 | — | — | — | — | — | |
| Pi0XPU Type=B60, Memory(GB)=24, Bandwidth(GB/s)=4562026.04 | — | 86 | 599 | 6.363 | 306.5 | — | — | — | — | — | |
| Pi0XPU Type=310p, Memory(GB)=48, Bandwidth(GB/s)=204.82026.04 | — | 86 | 1,030 | 2.618 | 818 | — | — | — | — | — |