Day 11 — Hand-as-support / loco-manipulation on the same ceilings

Day 10 owned reach and grasp as soft EE / hand tasks under measured foot contact_set, Day 07 friction / unilaterality, and torque_scale — hands stayed FREE unless measured contact honestly said otherwise; claiming SUPPORT without measurement was forbidden. Day 09 owned mid-event abort / soften. Day 08 owned gait modes vs measured contact. Day 07 owned the feasible program. Day 06 owned the estimate. Day 05 owned sag honesty.

Today owns the contact-mode change: promote measured hand contact into the Day 07 hard support set beside the feet — same friction cones, same $F_z \ge 0$, same torque_scale — so loco-manipulation (wall plant, table brace, handrail, crawl transition) is still one WBC/QP plant, not a bolted “arm balancer.”

Hand-as-support is contact mode, not a bigger EE weight

A palm on a wall is not “raise arm task weight until the QP tips the CoM onto the hand.” It is a support membership change: the hand enters measured contact_set, hard constraints attach, soft EE on that limb goes away or softens to residual contact regulation, and free limbs keep soft tasks under the expanded support polygon.

Core thesis: loco-manipulation consumes the same EstimatorState + Day 07 friction / unilaterality + torque_scale. Measured contact_set owns which hands and feet are SUPPORT. Soft EE / grasp remain for FREE limbs. There is still one QP face — not a second arm plant that steals wrench or finishes a wall plant into brownout.

Ownership rule: one writer for loco-manip / support claim intent and one writer for measured contact_set (feet and hands). When they disagree mid-plant, measured contact wins for constraints. Intent aborts, softens, or waits — it does not attach hand friction because a teleop stick said “lean now.”

How today fits the Days 05–10 stack

Do not bolt a Cartesian “arm balancer” that writes JointCommand beside the QP when a hand touches a surface. That is Day 10’s second-plant anti-pattern wearing a wall-plant demo.

Promoting measured hand into the support set

Day 10’s honesty table becomes an enable gate:

Promotion checklist (all required):

1. Intent allows CLAIM_SUPPORT or LOCO_MANIP (not a surprise bump).
2. Measured contact confidence above gate; contact age inside budget.
3. Hand wrench / normal agrees with expected surface (not a finger brush).
4. torque_scale and power path green enough to hold the new cone set.
5. QP remains feasible after attaching hand constraints — else abort claim, keep FREE.

Demotion is mandatory when any of those fail mid-contact: hand leaves hard set, soft EE may return only if intentionally FREE again, CoM bias retreats toward remaining measured support (usually feet).

Loco-manipulation under shared ceilings

Loco-manipulation here means using an expanded measured support set while still walking intent, bracing, or transferring weight — not a separate locomotion stack for arms.

Hard vs soft: support constraints on measured feet and hands stay hard. Soft EE authority lives only on FREE limbs. If the QP goes infeasible under the expanded set, the honest answer is demote / abort / soften — not silently dropping a hand cone so the teleop lean “makes it.”

Weight transfer honesty: moving CoM toward a hand is allowed only after MEASURED_SUPPORT. Scripted plants before free lean. Never free a still-loaded sole solely to clear a prettier palm pose.

Mid-contact abort / soften honesty

Hand-as-support is where teams disable safety “just until the lean settles.” Do not.

Abort, demote, and soften are success paths when the palm lied, the wall was grease, or the pack sagged. A robot that always finishes the planned lean into a collapsing bus is ignoring Day 05 with a prettier multi-contact demo.

Shared API: hand support on the same stack

Keep Day 02 / 04 wire honesty. Cyclic commands still carry BusHeader. Extend Day 10’s ManipulationCommand / WbcTaskSet and Day 06–08 ContactState — do not invent a parallel “arm balance 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 AND hands when measured — each SupportContact
  SupportContact:
    link_id / frame    # e.g. left_foot, right_palm
    kind               # FOOT | HAND
    mu, normal, wrench_est?
    confidence, age_ms
    flags              # settling / flicker / gated

# Thin loco-manip face (host / mid-rate) — extends Day 10
ManipulationCommand ⊇ BusHeader:
  intent               # IDLE | REACH | GRASP | HOLD | ABORT
                       # | CLAIM_SUPPORT | LOCO_MANIP
  ee_refs[]            # soft tasks — only meaningful for FREE limbs
  hand_refs[]          # soft grasp / residual contact regulation hints
  hand_role[]          # FREE | CLAIM_SUPPORT | MEASURED_SUPPORT (mirror of truth)
  support_claim[]      # which hand frames intent wants loaded — measured still owns truth
  priority_hint[]
  relax_on_fault       # stale / POWER_GATED / low torque_scale / flicker → abort/demote

# Extended Day 07–10 task face
WbcTaskSet ⊇ BusHeader:
  mode                 # STANCE_DS | STANCE_SS | SWING | IMPACT | RECOVER_IP
                       # | MANIP | HAND_SUPPORT | LOCO_MANIP | ...
  com_ref / capture_ref?
  foot_refs[]          # support | swing | settling
  hand_support_refs[]? # residual contact / unload bias for MEASURED_SUPPORT hands
  ee_refs[]            # soft EE — FREE limbs only
  hand_refs[]          # soft grasp — FREE hands only
  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 claim / loco-manip intent. Measured ContactState owns support membership. Rate loops enforce BMS ceilings. Nobody re-implements FOC or AHRS inside the wall-plant planner.

Physical ↔ digital pairing for hand-as-support / loco-manipulation

If the twin always plants with perfect Boolean palms and a full pack while hardware sees delayed contact_set and sag, you will tune lean gains forever and never ship a brace that survives a tired bus.

Bring-up sequence (before free loco-manip / teleop lean)

Days 07–10 green (stance, step, recovery, FREE-hand reach/grasp with mid-reach faults proven). Then:

1. Days 07–10 green — FREE-hand manipulation abort paths still behave; no open recovery or gait fights.
2. Quiet measured hand load — palm against a known surface at low normal; confirm contact enters contact_set with confidence; no CoM transfer yet; feet remain primary support.
3. One scripted plant — CLAIM → wait for MEASURED_SUPPORT → attach hand cones; soft EE on that hand drops to residual; confirm QP feasible.
4. Slow weight shift — scripted CoM bias toward expanded support; feet stay in measured set; monitor wrench share and torque_scale.
5. Mid-contact faults — force stale / gated / derate / hand flicker in twin and on hardware during plant and during lean; confirm demote / abort / soften, not heroic completion.
6. Then add teleop or learned residuals that choose claim / lean / loco-manip above the QP — never replace cones, FOC, or torque_scale. Multi-contact transitions and object-as-support wait for Day 12+.

Skip to “RL wall plants in sim” and hardware’s first real lean becomes an unplanned tip / brownout experiment with better video.

Failure gallery

What to freeze this week

Extend frames and your control config with:

• Hand support frames, normals, mu, and wrench / confidence gates aligned with feet
• Who owns support claim intent vs measured hand+foot contact_set (disagree → demote/hold)
• ContactState.SupportContact for HAND kind + age / flicker flags
• ManipulationCommand CLAIM_SUPPORT / LOCO_MANIP + hand_role / support_claim fields
• WbcTaskSet modes HAND_SUPPORT / LOCO_MANIP; soft EE only on FREE limbs
• Mid-contact abort / demote for stale, POWER_GATED, low torque_scale, and contact flicker
• Twin hooks: matched palm loads, contact delay/dropout, sag under brace peaks — same as hardware
• Bring-up gate: no free teleop lean / learned loco-manip until steps 1–5 above are green

The lesson is not “add more end-effector gains until the robot can lean.” It is make hand-as-support the same contact-rich program in silicon and in the twin: one estimate, honest power ceilings, measured hands in the hard support set beside feet, soft tasks only on FREE limbs, and abort / demote paths that stay honest when the palm or the pack lies mid-contact.

Next

Day 12 — Multi-contact transitions / object-as-support / whole-body loco-manip scheduling: sequenced promote–demote across feet, hands, and grasped objects; contact schedules that do not fight Day 08 gait or Day 09 recovery; still one plant, still measured contact owns the hard set.