Robotic Manipulation on Simulation Long-Horizon
67.08Success RateBest Hierarchy
Evaluation results
| Method | Links | |
|---|---|---|
| Best HierarchyConfiguration=Combination of best-performing parameters: cross-episodic memory, thinking VLM, contact-based observation, success-based termination2026.06 | 67.08 | |
| Naive HierarchyConfiguration=Combination of naively-chosen parameters: no memory, VLM without thinking, raw observation, fixed-horizon termination2026.06 | 40.56 | |
| Flat VLAConfiguration=No VLM planner; task prompt directly passed to low-level VLA2026.06 | 25.3 |