Resolve the contact-only LEVER task

Establish reliable end-to-end completion of the two-step contact-only CPU socket LEVER closure task, which remains unresolved despite the value-guided refinement and local adaptation components of Facet-0.

Background

The LEVER benchmark is the only task in the five-task suite without a free-space sub-goal; it requires two sequential presses to close the CPU socket lever. Facet-0 reaches only 50% task success on this benchmark, and the intermediate +RL variant reaches 0%, showing that the remaining difficulty is concentrated in contact-critical pressing rather than approach or grasping.

The authors explicitly identify the contact-only LEVER task as unresolved. This indicates an unresolved reliability problem in handling consecutive force-sensitive presses, even though the full system performs strongly on the four insertion tasks.

References

LEVER remains the hardest at 50\% because both commands are contact-critical presses and either failure invalidates the trial.

Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation  (2609.01596 - Deng et al., 1 Sep 2026) in Section 4.1, “Main Comparison and Component Gains (Q1)”