E8 / Post-alarm exit actions

01 / QUESTION

Fifteen fixed-scene rollouts compare hold, open and open-then-retreat after identical alarms.

METHOD

Five conditions × hold/open/open-retreat: 15 fixed-scene rollouts.

RESULT

Opening removes final residual grasp in late lowering/silence; all transfer-drop policies still fail.

OBSERVATIONS

Fifteen pre-alarm state pairs match; open and open-retreat have identical final acceptance.

CONCLUSION

Identical alarms do not imply identical physical outcomes; exit actions need independent checks.

LIMITATIONS

One deterministic rollout per cell, known tray, no force/hardware safety validation; falling into the tray is not planned release.

REPRODUCE

See the pinned repository for code, frozen parameters, raw results and independent auditing.

1. Engineering cases

   Embodied AI VIII: what should the gripper do after a drop alarm?

   Fifteen fixed-scene rollouts compare hold, open and open-then-retreat after identical alarms.

   2026-10-039 MINPublished
2. PROJECT / 003

   Hohoo Embodied Experiments

   Two MuJoCo teaching scenes—pick-and-place and planar force control—record contracts, tuning, clearance and delayed-feedback outcomes, including failures.

   2026-09-28Experiment