List your company

Robotic task execution on LIBERO

98.4Average Success Rateπ0.5-ADV
1.78426.86751.9577.033Mar 18, 2026Mar 29, 2026Apr 10, 2026Apr 22, 2026May 4, 2026May 16, 2026May 28, 2026

Evaluation results

MethodLinks
π0.5-ADVAction Chunk=16, Steps per Second=37.3hz2026.0398.498.499.498.297.6———
InternVL3.5-ADVAction Chunk=16, Steps per Second=36.9hz2026.0397.897.699.497.296.8———
π0.5 + GridS16Vis. Tokens=16, FLOPs (G)=83.84, Time (s)=6.762026.0597.798.698.898.495.21——
Qwen2.5-VL-ADVAction Chunk=16, Steps per Second=41.0hz2026.0396.896.698.89794.6———
π0.5Vis. Tokens=256, FLOPs (G)=249.76, Time (s)=8.542026.0596.798.49897.692.8———
π0.5 + GridS4Vis. Tokens=4, FLOPs (G)=75.73, Time (s)=6.532026.0596.797.49997.492.80——
π0.5Action Chunk=16, Steps per Second=61.6hz2026.0396.597.898.496.693———
π0 + GridS16Vis. Tokens=16, FLOPs (G)=51.65, Time (s)=6.052026.05969899.296.490.21.6——
OpenVLA-OFT (STRONG)Perturbation Modality=N/A, Perturbation Family=N/A, Perturbation Type=Clean, Zero-shot robustness=false2026.0495.75————-0.75——
OpenVLA-OFT (STRONG)Perturbation Modality=Visual, Perturbation Family=Photometric Degradation, Perturbation Type=Color jitter, Zero-shot robustness=false2026.0495.75————1.25——
π0 + GridS4Vis. Tokens=4, FLOPs (G)=43.61, Time (s)=5.862026.0595.596.699.496.489.61.1——
OpenVLA-OFTVis. Tokens=512, FLOPs (G)=3922.56, Time (s)=17.432026.0594.897.694.295.292———
InternVL3.5-DiffusionAction Chunk=16, Steps per Second=57.6hz2026.0394.795.89694.292.8———
π0Vis. Tokens=256, FLOPs (G)=216.01, Time (s)=8.172026.0594.497.298.89685.6———
Qwen2.5-VL-DiffusionAction Chunk=16, Steps per Second=68.4hz2026.0393.994.697.293.890.2———
π0 + VLA-CacheVis. Tokens=256, FLOPs (G)=188.12, Time (s)=7.52, training-free=true2026.0593.895.297.696.685.6-0.6——
CogACTVis. Tokens=256, FLOPs (G)=1891.24, Time (s)=13.802026.0593.697.29890.288.8———
OpenVLA-OFT (STRONG)Perturbation Modality=Textual, Perturbation Family=Semantics Drift, Perturbation Type=Spatial, Zero-shot robustness=true2026.0493.5————0——
π0 + FastVVis. Tokens=100, FLOPs (G)=143.59, Time (s)=7.32, training-free=true2026.0592.99798.493.882.4-1.5——
π_0 (STRONG)Perturbation Modality=N/A, Perturbation Family=N/A, Perturbation Type=Clean, Zero-shot robustness=false2026.0492.5————2.75——
π_0 (STRONG)Perturbation Modality=Textual, Perturbation Family=Semantics Drift, Perturbation Type=Spatial, Zero-shot robustness=true2026.0492.25————0.75——
π_0 (STRONG)Perturbation Modality=Visual, Perturbation Family=Photometric Degradation, Perturbation Type=Color jitter, Zero-shot robustness=false2026.0490.75————-0.75——
π_0 (STRONG)Perturbation Modality=Textual, Perturbation Family=Adversarial Injection, Perturbation Type=Suffix, Zero-shot robustness=false2026.0490.25————21.5——
π0 + SparseVLMVis. Tokens=100, FLOPs (G)=150.30, Time (s)=7.48, training-free=true2026.0589.893.49891.276.6-4.6——
InternVL3.5-FASTAction Chunk=16, Steps per Second=3.2hz2026.0389.689.692.893.682.8———
OpenVLA-OFT (STRONG)Perturbation Modality=Visual, Perturbation Family=Geometric Shifts, Perturbation Type=Random crop, Zero-shot robustness=false2026.0489.5————7——
Qwen2.5-VL-FASTAction Chunk=16, Steps per Second=5.0hz2026.0387.990.288.696.276.4———
π_0 (STRONG)Perturbation Modality=Multimodal, Perturbation Family=N/A, Perturbation Type=Prefix adversarial injection + Color jitter, Zero-shot robustness=false2026.0487.5————39.75——
π_0 (STRONG)Perturbation Modality=Visual, Perturbation Family=Geometric Shifts, Perturbation Type=Random crop, Zero-shot robustness=false2026.0483.75————11——
Oracle abstention (upper bound)epsilon=0.05, Number of cells=2502026.0583—————091
SmolVLAVis. Tokens=64, FLOPs (G)=306.78, Time (s)=8.912026.057981.392.985.855.8———
OpenVLAVis. Tokens=256, FLOPs (G)=1956.19, Time (s)=26.082026.0576.584.788.479.253.7———
OpenVLA-OFT (STRONG)Perturbation Modality=Textual, Perturbation Family=Adversarial Injection, Perturbation Type=Suffix, Zero-shot robustness=false2026.0472.25————70.5——
BOKBO globalepsilon=0.05, Number of cells=2502026.0571—————3.180
OpenVLA (STRONG)Perturbation Modality=N/A, Perturbation Family=N/A, Perturbation Type=Clean, Zero-shot robustness=false2026.0469.25————-1——
Held-out threshold (no guarantee)epsilon=0.05, Number of cells=2502026.0569—————3.479
BOKBO Mondrianepsilon=0.05, Number of cells=2502026.0569—————3.474
OpenVLA (STRONG)Perturbation Modality=Textual, Perturbation Family=Semantics Drift, Perturbation Type=Spatial, Zero-shot robustness=true2026.0468.25————3.75——
OpenVLA (STRONG)Perturbation Modality=Visual, Perturbation Family=Photometric Degradation, Perturbation Type=Color jitter, Zero-shot robustness=false2026.0468.25————1.75——
No abstention (K-sample only)epsilon=0.05, Number of cells=2502026.0568—————8.6100
OpenVLA (STRONG)Perturbation Modality=Multimodal, Perturbation Family=N/A, Perturbation Type=Prefix adversarial injection + Color jitter, Zero-shot robustness=false2026.0466.25————9.5——
OpenVLA (STRONG)Perturbation Modality=Textual, Perturbation Family=Adversarial Injection, Perturbation Type=Suffix, Zero-shot robustness=false2026.0466————30——
OpenVLA-OFT (STRONG)Perturbation Modality=Multimodal, Perturbation Family=N/A, Perturbation Type=Prefix adversarial injection + Color jitter, Zero-shot robustness=false2026.0465.75————41.5——
OpenVLA (STRONG)Perturbation Modality=Visual, Perturbation Family=Geometric Shifts, Perturbation Type=Random crop, Zero-shot robustness=false2026.045.5————0.25——