E7 / Grasp contracts during lowering

01 / QUESTION

What changes when transfer monitoring extends into lowering?

METHOD

Seven conditions × three policies: 21 deterministic rollouts isolating monitoring scope and predicates, with a common latched target hold and no recovery or replanning.

RESULT

The reused height rule falsely stops clean lowering at 6.002 s. Stage-aware checks allow clean placement and alarm 40 ms after the first post-fault capture in both lowering-opening cases.

OBSERVATIONS

All early-drop runs pass the endpoint check. Late stage-aware stopping leaves both fingers touching the cube and fails placement. Silence and stale replay trigger the 60 ms age limit.

CONCLUSION

Stage determines predicate meaning. A correct alarm does not validate the exit action; passing the endpoint does not validate the process.

LIMITATIONS

One deterministic run per cell, fixed phases, one clock and privileged simulated observations; no contact-force analysis, post-release contract, vision, ROS2 or hardware validation.

REPRODUCE

21 tests pass. Independent auditing covers 105 episode files, 10 source fingerprints and 21 pre-decision prefixes. Raw CSV, state snapshots and figures are archived.

1. Engineering cases

   Embodied AI Practice VII: Why a Transfer Rule Stops Normal Lowering

   21 MuJoCo rollouts separate stage-specific grasp checks, stop actions and final placement. A false alarm and lingering finger contact reveal two different failure modes.

   2026-10-0113 MINPublished
2. Engineering cases

   Embodied AI VI: what do recovery gating and replanning each solve?

   A 30-rollout ablation reduces first target jumps from 50–97 mm to about 0.25 mm, without improving placement outcomes in this fixed matrix. Separate recovery evidence, command continuity and task success.

   2026-09-3014 MINPublished
3. LAB / 012

   E6 / Recovery gate and path ablation

   Separate revalidation from replanning: which differences concern evidence and which concern the resumed target?

   2026-09-30Completed
4. PROJECT / 003

   Hohoo Embodied Experiments

   Seven MuJoCo teaching experiments, from pickup bias to stage contracts, with raw trajectories, decisions and independent audits.

   2026-09-28Experiment