Robotic manipulation on Fractal
90.82Success RateVISUALTHINK-VLA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VISUALTHINK-VLAReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 90.82 | 0.367 | |
| FULLSOFTReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 90.38 | 0.448 | |
| BASEVLAReasoning Strategy=OpenVLA-family re-evaluation2026.05 | 87.45 | 0.345 | |
| TraceVLAReasoning Strategy=Spatial or image-grounded thinking2026.05 | 86.25 | 0.402 | |
| SIEVEBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=50K2026.07 | 76.4 | — | |
| Full-TrainingBase Model=Qwen3-VL-4B-GR00T, Selection Budget=100%, Training Steps=100K2026.07 | 75 | — | |
| SCIZORBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=50K2026.07 | 71.9 | — | |
| SpatialVLAReasoning Strategy=Spatial or image-grounded thinking2026.05 | 67.85 | 0.582 | |
| DemInfBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=50K2026.07 | 67.4 | — | |
| RandomBase Model=Qwen3-VL-4B-GR00T, Selection Budget=50%, Training Steps=50K2026.07 | 55.6 | — | |
| DeepThinkVLA-RLReasoning Strategy=Reinforced or hybrid thinking-action decoding2026.05 | 33.67 | 1.219 | |
| ECoTReasoning Strategy=Textual CoT / autoregressive reasoning before action2026.05 | 32.87 | 6.702 | |
| InternVLA-M1Reasoning Strategy=Spatial or image-grounded thinking2026.05 | 23.46 | 1.153 |