Robotic Manipulation on BridgeData V2
89.49Success RateVISUALTHINK-VLA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VISUALTHINK-VLAReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 89.49 | 0.367 | |
| FULLSOFTReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 88.45 | 0.447 | |
| TraceVLAReasoning Strategy=Spatial or image-grounded thinking2026.05 | 86.87 | 0.404 | |
| SpatialVLAReasoning Strategy=Spatial or image-grounded thinking2026.05 | 86.57 | 0.589 | |
| ECoTReasoning Strategy=Textual CoT / autoregressive reasoning before action2026.05 | 85.09 | 8.377 | |
| BASEVLAReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 75.37 | 0.345 | |
| SIEVEBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=25K2026.07 | 56.3 | — | |
| SCIZORBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=25K2026.07 | 52.2 | — | |
| Full-TrainingBase Model=Qwen3-VL-4B-GR00T, Selection Budget=100%, Training Steps=50K2026.07 | 51.8 | — | |
| DemInfBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=25K2026.07 | 43.2 | — | |
| RandomBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=25K2026.07 | 39.6 | — | |
| DeepThinkVLA-RLReasoning Strategy=Reinforced or hybrid thinking-action decoding2026.05 | 27.93 | 1.204 | |
| InternVLA-M1Reasoning Strategy=Spatial or image-grounded thinking2026.05 | 11.18 | 1.153 |