Day 07 — Contact-rich balance policies & whole-body control

Day 06 gave you one balance-ready EstimatorState on a shared clock, with ages, contact modes, and power honesty. Day 05 already made torque_scale a control surface. Day 02 made $i_q$ the muscle.

Today owns the consumer: how contact-rich balance and whole-body control (WBC) turn that estimate into JointCommand without inventing a second plant — and without stuffing residual physics into a black-box net that runs inside a microsecond loop.

Policy consumes estimates; it does not re-fuse them

Balance and WBC answer a different question than the estimator:

Rule: one writer for fusion (Day 06). Many readers. If teleop, WBC, and a learned policy each run their own AHRS or “quick CoM fix,” twins and hardware diverge on day one.

Treat INPUT_STALE, POWER_GATED, low cov_or_confidence, and empty / wrong contact_set the way Day 05 treats brownout: behavior must change. Soften peaks, freeze aggressive tasks, widen friction margins, or open a safe torque path — do not hope the QP will “average it out.”

Tasks vs constraints (and where power sits)

A practical WBC / QP sketch (hierarchical or weighted — the contract matters more than the brand):

Put torque_scale on the constraint / ceiling side, not as a silent gain inside a learned residual. When the pack sags, the feasible set shrinks; the policy that pretends otherwise only balances on a full pack.

Neural nets: fine for gait mode, footstep choice, or residual correction above measured-jitter rate loops. The site principle still holds — never put a net inside a microsecond current or torque loop until you have measured jitter and proven the plant contract. Prefer nets that read the same EstimatorState the QP reads.

Contact-rich: the support story changes the program

Contact events are sparse; the QP’s geometry is not optional flavor text.

Physical and twin must agree on when contact_set flips relative to $F_z$ / CoP (Day 06 exercise). A policy that learned Boolean contacts from a perfect physics engine will heel-strike late on hardware every time.

Shared API: command the muscle honestly

Keep Day 02 / 04 wire honesty. Cyclic commands still carry BusHeader. A thin balance / WBC face looks like this:

BusHeader:
  node_id, seq
  t_sample, clock_domain, age_budget_ms
  faults

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

# Optional task face (host / mid-rate)
WbcTaskSet ⊇ BusHeader:
  mode                 # STANCE_DS | STANCE_SS | STEP | ...
  com_ref / cop_ref?   # in named frames from frames contract
  foot_refs[]          # support vs swing tagged
  posture_bias?
  priority_order[]     # or weights — pick one story and freeze it
  relax_on_fault       # what to drop when POWER_GATED / INPUT_STALE

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

Higher layers should not re-implement FOC or AHRS. They publish tasks or setpoints; the rate loops enforce ceilings the BMS already declared.

Physical ↔ digital pairing for balance / WBC

Bring-up sequence (before trusting closed-loop balance)

Day 06 steps 1–6 stay green. Then:

1. Gravity compensate / hold — joint-level honesty with estimator watched, balance tasks off or weight≈0.
2. Quiet double support — small posture + CoM regulation; plot CoP vs polygon and contact_set.
3. Slow weight shift — CoP travels; solver stays feasible; no swing-foot constraints left on.
4. Single-support unload — mode switch matches hardware; swing foot free.
5. Fault injection — force stale / POWER_GATED / low torque_scale in twin and on hardware; confirm soften / freeze paths.
6. Then add steps or learned residuals that consume the same state — never replace the QP’s friction and power ceilings.

Skip to “just stand with the net” and you will tune “balance gains” while fighting a flipped contact bit or an infinite bus in sim.

Failure gallery

What to freeze this week

Extend frames and your control config with:

• Who may write JointCommand during stance (WBC only vs teleop overlay rules)
• Task list + priority / weight story for DS vs SS (and impact soften)
• How torque_scale / PowerFault maps to tau or abs $i_q$ ceilings every tick
• Fallback when QP is infeasible or EstimatorState is stale / gated
• Twin hooks: delay, contact timing, sag — same budgets as hardware
• Bring-up gate: no free stepping / learned residuals until steps 1–5 above are green

The lesson is not “pick OSQP vs a hierarchy.” It is make contact-rich balance the same feasible program in silicon and in the twin, reading one estimate, honoring power, and keeping black-box residuals above the loops you can measure.

Next

Day 08 — Contact-rich stepping & locomotion: footstep timing, swing trajectories, and gait modes that still consume EstimatorState + WBC ceilings instead of bypassing them.