Pick-Place on AgileX PiPER real-world
93.3Success RateLoRA-SP
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LoRA-SPModel=SmolVLA, Trainable / Total (%)=17.1, Active Rank=60, Training regime=Multi-task2026.03 | 93.3 | — | |
| MotionVLA2026.06 | 92.5 | 119 | |
| Full FTModel=π0, Trainable / Total (%)=100.0, Active Rank=Full, Training regime=Multi-task2026.03 | 86.7 | — | |
| LoRA (r = 128)Model=SmolVLA, Trainable / Total (%)=17.0, Active Rank=128, Training regime=Multi-task2026.03 | 86.7 | — | |
| Full FTModel=SmolVLA, Trainable / Total (%)=100.0, Active Rank=Full, Training regime=Multi-task2026.03 | 86.7 | — | |
| LoRA-MoE (weighted sum)Model=π0, Trainable / Total (%)=9.2, Active Rank=128, Training regime=Multi-task2026.03 | 80 | — | |
| LoRA-SPModel=π0, Trainable / Total (%)=9.2, Active Rank=76, Training regime=Multi-task2026.03 | 80 | — | |
| π02026.06 | 80 | 118 | |
| LoRA-MoE (top-1)Model=SmolVLA, Trainable / Total (%)=17.2, Active Rank=32, Training regime=Multi-task2026.03 | 73.3 | — | |
| LoRA-MoE (weighted sum)Model=SmolVLA, Trainable / Total (%)=17.2, Active Rank=128, Training regime=Multi-task2026.03 | 66.7 | — | |
| LoRA (r = 128)Model=π0, Trainable / Total (%)=9.1, Active Rank=128, Training regime=Multi-task2026.03 | 60 | — | |
| AdaLoRAModel=π0, Trainable / Total (%)=9.1, Active Rank=76, Training regime=Multi-task2026.03 | 60 | — | |
| AdaLoRAModel=SmolVLA, Trainable / Total (%)=17.0, Active Rank=60, Training regime=Multi-task2026.03 | 20 | — | |
| BaselinePolicy architecture=pi0.5, Method variant=Baseline2026.03 | 20 | — | |
| DepthCachePolicy architecture=pi0.5, Method variant=+ Ours2026.03 | 20 | — | |
| LoRA-MoE (top-1)Model=π0, Trainable / Total (%)=9.2, Active Rank=32, Training regime=Multi-task2026.03 | 13.3 | — | |
| Demo2026.06 | — | 119 |