Robotic Manipulation on Real-robot manipulation tasks Aggregate
95Average Success Rate (Avg SR)π0.5
Evaluation results
| Method | Links | ||
|---|---|---|---|
| π0.5Execution mode=synchronous2026.06 | 95 | — | |
| M2-ResiPolicy2026.03 | 94 | — | |
| UniVLA-AsyncExecution mode=asynchronous2026.06 | 93.3 | — | |
| RDPsensor=Tactile2026.03 | 87 | — | |
| RDPsensor=Force2026.03 | 82 | — | |
| UniVLAExecution mode=synchronous2026.06 | 78.3 | — | |
| LaST-HDTraining Data=100 real-robot demonstrations, Rollouts=20 per task2026.06 | 73 | — | |
| CFM-StudentInference Steps=1-step2026.03 | 70 | 125 | |
| LaST-HD (Mix-HD)Training Data=50 robot and 50 OOL Glove demonstrations, Rollouts=20 per task2026.06 | 68 | — | |
| DPtactile=true2026.03 | 65 | — | |
| SmolVLA2026.06 | 65 | — | |
| LaST0Training Data=100 real-robot demonstrations, Rollouts=20 per task2026.06 | 63 | — | |
| π0.5Training Data=100 real-robot demonstrations, Rollouts=20 per task2026.06 | 62 | — | |
| DPbaseline=true2026.03 | 55 | — | |
| CFM-TeacherInference Steps=50-step2026.03 | 54 | 2.9 | |
| Cosmos-PolicyTraining Data=100 real-robot demonstrations, Rollouts=20 per task2026.06 | 52 | — | |
| PointFlowMatchInference Steps=50-step2026.03 | 48.7 | 3.6 | |
| A2C22026.06 | 40 | — | |
| PointFlowMatchInference Steps=1-step2026.03 | 3.3 | 70 |