Day 12 — Multi-contact transitions / object-as-support / loco-manip scheduling

Day 11 owned promoting measured hand contact into the Day 07 hard support set beside the feet — CLAIM without measure forbidden, mid-contact demote on flicker / stale / POWER_GATED, still one plant. Day 10 owned soft EE / grasp on FREE limbs. Day 09 owned mid-event abort and RecoveryCommand. Day 08 owned gait modes vs measured contact_set. Day 07 owned the feasible program. Day 06 owned the estimate. Day 05 owned sag honesty.

Today owns the contact schedule: sequenced promote–demote across feet, hands, and grasped objects (object-as-support when measured), without a second plant and without fighting Day 08 gait clocks or Day 09 recovery. Schedules are intent that wait for measurement. Measured contact still owns the hard set.

Contact schedule is ownership of transitions, not a prettier constraint clock

A multi-contact “choreography” that attaches hand and object cones because the schedule clock said $t = 1.2\,\mathrm{s}$ is just CLAIM-as-truth with more limbs. The schedule proposes who should take load next; measurement admits them; demotion is mandatory when confidence, wrench agreement, or ceilings fail mid-transition.

Core thesis: multi-contact transitions consume the same EstimatorState + Day 07 friction / unilaterality + torque_scale. A ContactSchedule (or thin MultiContactCommand) sequences CLAIM → wait-for-measure → PROMOTE → HOLD → DEMOTE across limbs and grasped objects. SupportContact.kind = OBJECT joins the hard set only when measured and wrench-agreed. Soft EE remains for FREE limbs. There is still one QP face — not a schedule-driven second plant that finishes a crate brace into a collapsing bus.

Ownership rule: one writer for schedule / multi-contact intent and one writer for measured contact_set. When they disagree mid-transition, measured contact wins for constraints. Intent aborts, softens, retimes, or waits — it does not attach OBJECT cones because the choreography said “lean on the crate now.” Day 08 gait and Day 09 recovery outrank a stuck schedule hold.

How today fits the Days 05–11 stack

Do not bolt a “multi-contact optimizer” that writes JointCommand beside the QP when the schedule advances. That is Day 10’s second-plant anti-pattern wearing a crate-brace demo.

Promote–demote sequencing across feet, hands, and objects

Day 11’s hand honesty table generalizes to a per-slot lifecycle:

Sequencing rules (all required):

1. One promote at a time (or a documented simultaneous pair) — no “attach every cone the schedule listed this tick.”
2. Intent CLAIMs the next slot; measurement must confirm before cones attach.
3. Demote the unloading contact only after the loading contact is MEASURED and the QP stays feasible — never free a still-loaded sole or palm to clear a prettier pose.
4. OBJECT slots require grasp integrity and object–world wrench agreement (not just “fingers closed”).
5. Mid-sequence POWER_GATED / stale / recovery → freeze or abort the schedule; do not advance phases.

Demotion is success when the crate slipped, the palm flickered, or the pack sagged. A robot that always finishes the next schedule beat into a collapsing bus is ignoring Day 05 with a prettier multi-contact demo.

Object-as-support honesty

A grasped crate, handrail, or table edge is not an end-effector weight. It is a candidate SupportContact with kind = OBJECT that joins the hard set only when:

Forbidden: treating “we are holding something” as object-as-support; attaching object cones from grasp alone; using a second Cartesian plant to “balance through the crate”; promoting OBJECT while Day 09 recovery owns the mode.

Object frames, mu, and normals belong in the same documented cone family as feet and hands (Day 07 honesty). If you cannot name the contact frame and friction model, you are not ready to promote OBJECT.

Schedule vs gait / recovery — non-fight rules

Multi-contact scheduling is where teams accidentally create three writers on WbcTaskSet.mode. Freeze a priority:

Hard vs soft: support constraints on measured FOOT / HAND / OBJECT stay hard. Soft EE authority lives only on FREE limbs. If the QP goes infeasible under the expanded set, demote / abort / soften — do not silently drop an object cone so the choreography “makes it.”

Mid-transition abort / soften honesty

Transitions are where teams disable safety “just until the handoff settles.” Do not.

Abort, demote, and soften are success paths when the object lied, the wall was grease, or the pack sagged mid-handoff.

Shared API: schedule + OBJECT on the same stack

Keep Day 02 / 04 wire honesty. Cyclic commands still carry BusHeader. Extend Day 10–11 ManipulationCommand / WbcTaskSet and Day 06–08 ContactState — do not invent a parallel “multi-contact bus”:

BusHeader:
  node_id, seq
  t_sample, clock_domain, age_budget_ms
  faults

# Upstream (read-only here)
EstimatorState, PowerState  ⊇ BusHeader

ContactState ⊇ BusHeader:
  contact_set[]        # feet, hands, AND objects when measured
  SupportContact:
    link_id / frame    # e.g. left_foot, right_palm, crate_top
    kind               # FOOT | HAND | OBJECT
    mu, normal, wrench_est?
    confidence, age_ms
    flags              # settling / flicker / gated / grasp_lost
    parent_grasp_id?   # for OBJECT: which grasp frames this support

# Thin multi-contact face (host / mid-rate)
ContactSchedule ⊇ BusHeader:          # or MultiContactCommand
  intent               # IDLE | MULTI_CONTACT | OBJECT_SUPPORT | ABORT
  slots[]:
    target_frame       # FOOT / HAND / OBJECT frame
    phase              # FREE | CLAIM | PROMOTE | HOLD | DEMOTE
    earliest_t / latest_t?   # soft timing hints — NOT constraint truth
    priority_hint
  yield_to             # GAIT | RECOVERY | POWER   # explicit non-fight
  relax_on_fault       # stale / POWER_GATED / flicker / grasp_lost → abort/demote

# Extends Day 10–11 manipulation face (optional merge)
ManipulationCommand ⊇ BusHeader:
  intent               # … | CLAIM_SUPPORT | LOCO_MANIP
                       # | OBJECT_SUPPORT | MULTI_CONTACT | ABORT
  ee_refs[] / hand_refs[]
  hand_role[] / support_claim[]
  object_claim[]       # which OBJECT frames intent wants loaded
  schedule_ref?        # link to active ContactSchedule id
  relax_on_fault

# Extended Day 07–11 task face
WbcTaskSet ⊇ BusHeader:
  mode                 # STANCE_DS | STANCE_SS | SWING | IMPACT | RECOVER_IP
                       # | MANIP | HAND_SUPPORT | LOCO_MANIP
                       # | OBJECT_SUPPORT | MULTI_CONTACT | ...
  com_ref / capture_ref?
  foot_refs[]
  hand_support_refs[]?
  object_support_refs[]?  # residual contact / unload bias for MEASURED OBJECT
  ee_refs[]               # soft EE — FREE limbs only
  hand_refs[]
  posture_bias?
  priority_order[]
  relax_on_fault

# Downstream muscle (Day 02) — ceilings still reflect torque_scale
JointCommand ⊇ BusHeader:
  q_des? / qd_des? / tau_ff?
  iq_ceiling? / tau_ceiling?
  flags                # STO / soften / hold

Higher layers publish schedule / object-support intent. Measured ContactState owns support membership. Gait and recovery keep their writers. Rate loops enforce BMS ceilings. Nobody re-implements FOC or AHRS inside the multi-contact choreographer.

Physical ↔ digital pairing for multi-contact transitions

If the twin always promotes on the schedule clock with sticky OBJECT contacts and a full pack while hardware sees delayed contact_set and sag, you will tune handoff gains forever and never ship a transition that survives a tired bus.

Bring-up sequence (before free multi-contact choreography)

Days 07–11 green (stance, step, recovery, FREE-hand reach, measured hand SUPPORT / loco-manip with mid-contact faults proven). Then:

1. Days 07–11 green — hand SUPPORT abort paths still behave; no open recovery or gait fights.
2. Single-slot script — CLAIM → wait for MEASURED → PROMOTE → HOLD → DEMOTE on one known surface or fixed object; feet remain primary; confirm QP feasible.
3. Two-slot handoff — scripted promote of B only after A is healthy; demote A only after B holds load share; monitor wrench share and torque_scale.
4. OBJECT-as-support — grasped object with documented frame/mu; confirm grasp_lost and wrench-disagree demote; no CoM bias onto OBJECT until MEASURED.
5. Non-fight + mid-transition faults — force schedule-during-swing, recovery-during-HOLD, stale / gated / flicker / grasp pop in twin and on hardware; confirm yield / abort / demote, not heroic completion.
6. Then add teleop or learned residuals that choose schedule slots / timing above the QP — never replace cones, FOC, torque_scale, gait, or recovery writers. Perception-driven affordances and richer schedule discovery wait for Day 13+.

Skip to “RL multi-contact dances in sim” and hardware’s first real crate handoff becomes an unplanned tip / brownout experiment with better video.

Failure gallery

What to freeze this week

Extend frames and your control config with:

• OBJECT support frames, normals, mu, grasp_id linkage, and wrench / confidence gates aligned with feet and hands
• Who owns ContactSchedule intent vs measured contact_set vs gait vs recovery (disagree → wait/yield/demote; recovery and gait outrank schedule holds)
• SupportContact.kind = OBJECT + grasp_lost / flicker flags
• ContactSchedule / MultiContactCommand phases FREE → CLAIM → PROMOTE → HOLD → DEMOTE with yield_to
• WbcTaskSet modes OBJECT_SUPPORT / MULTI_CONTACT; soft EE only on FREE limbs
• Mid-transition abort / demote for stale, POWER_GATED, low torque_scale, flicker, and grasp integrity loss
• Twin hooks: matched object loads, contact delay/dropout, grasp pop, schedule-vs-gait collisions, sag under handoff peaks — same as hardware
• Bring-up gate: no free multi-contact choreography / learned schedules until steps 1–5 above are green

The lesson is not “play a contact MIDI file until the robot looks busy.” It is make multi-contact transitions the same contact-rich program in silicon and in the twin: one estimate, honest power ceilings, measured FOOT/HAND/OBJECT in the hard support set, schedules that wait for measurement and yield to gait/recovery, soft tasks only on FREE limbs, and abort / demote paths that stay honest when the object or the pack lies mid-handoff.

Next

Day 13 — Perception-driven support affordances / contact discovery feeding schedules: vision and tactile proposals become schedule candidates only; measured contact still owns the hard set; no detector-as-constraint-truth; affordance confidence gates CLAIM, not cones.