Day 08 — Contact-rich stepping & locomotion

Day 07 made contact-rich balance a feasible program: tasks vs constraints, friction / unilaterality, and torque_scale as a ceiling — not a silent gain. Day 06 owned the estimate. Day 05 owned sag honesty.

Today owns motion that still respects that program: footstep timing, swing trajectories, and gait modes that consume EstimatorState + WBC ceilings instead of inventing a second plant that “just walks.”

Locomotion is a mode machine on top of stance

Walking is not “balance with a sine wave on the feet.” It is a sparse schedule of contact geometries the QP already knows how to treat:

Ownership rule: pick one writer for intended mode (gait phase clock / GaitCommand) and one writer for measured support (contact_set from Day 06). The swing foot must not receive hard support constraints while it is still in contact_set, and the support foot must not be “freed” because a clock said so if the sole is still loaded. When they disagree, measured contact wins for constraints; the gait layer aborts, holds, or re-plans.

If teleop, a learned gait net, and a scripted clock each flip mode independently, the QP spends heel-strike fighting a foot that is still a support — or leaning into air.

Footstep timing, capture, and honest abort

A foothold is a promise about future support geometry. Capture-point / Divergent Component of Motion intuition still applies at the planning layer:

• Desired next contact should be reachable given CoM / capture-point state in EstimatorState, not a video-game waypoint.
• Step duration and length shrink when cov_or_confidence is poor, IMU / FT ages are near budget, or torque_scale is already derating the stance leg.
• Abort and shorten are first-class behaviors, not shame states. A robot that always completes a planned swing into a sagging bus is not “brave” — it is ignoring Day 05.

Do not let a high-level net “push through” a step the ceilings already declared infeasible. The net may choose a shorter step or a mode change; it may not replace friction cones or FOC.

Swing as soft tasks + clearance; land into Day 07 constraints

Swing is a task, not a second constraint world:

Touchdown: kinematic sole height crossing zero is a hint, not a contact event. Gate landing with the same ContactState / impact policy Day 06 froze. Then hand the new support to Day 07’s friction cone and $F_z \ge 0$ — do not keep treating the foot as a free end-effector after the sole is loaded.

RL / learned residuals: choose footsteps, cadence, or residual CoM bias above the QP. Never train a net that outputs joint torque while pretending cones and torque_scale do not exist. Prefer policies that read the same EstimatorState and fault bits the gait layer reads.

Shared API: thin gait face, honest muscle

Keep Day 02 / 04 wire honesty. Cyclic commands still carry BusHeader. Extend the Day 07 task face rather than inventing a parallel “walk bus”:

BusHeader:
  node_id, seq
  t_sample, clock_domain, age_budget_ms
  faults

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

# Thin locomotion face (host / mid-rate)
GaitCommand ⊇ BusHeader:
  mode                 # DS | SS_L | SS_R | STEP | ABORT | HOLD
  phase_or_clock?      # who owns intended timing — document it
  cadence_hz?
  relax_on_fault       # stale / POWER_GATED / low torque_scale → abort/shorten

FootstepPlan ⊇ BusHeader:
  footholds[]          # frame + pose + earliest/latest touchdown window
  swing_clearance?
  capture_ref?         # optional planning hint in named frames
  abort_policy         # shorten | freeze_ds | recover

# Extended Day 07 task face
WbcTaskSet ⊇ BusHeader:
  mode                 # STANCE_DS | STANCE_SS | SWING | IMPACT | ...
  com_ref / capture_ref?
  foot_refs[]          # each tagged support | swing | settling
  posture_bias?
  priority_order[]     # or weights — same story as Day 07
  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 mode + footholds + tasks. Rate loops enforce BMS ceilings. Nobody re-implements FOC or AHRS inside the swing interpolator.

Physical ↔ digital pairing for stepping

If the twin walks with perfect Boolean contacts and a full pack while hardware sees delayed contact_set and sag, you will tune swing gains forever and never ship a step.

Bring-up sequence (before free walking)

Day 07 steps 1–5 stay green (gravity hold → quiet DS → weight shift → controlled SS unload → fault injection). Then:

1. Stance green — DS/SS unload, friction, and ceilings behave under Day 07 fault injection.
2. Weight-shift only — CoP travels; no swing tasks; confirm contact_set tracking.
3. One controlled step — single FootstepPlan; mode flip tied to measured contact; land into cones.
4. Short cadence — a few steps at low speed; plot abort/shorten paths even if unused.
5. Mid-step faults — force stale / gated / derate in twin and on hardware during swing; confirm abort or soften, not a heroic completion.
6. Then add learned footstep / mode residuals that consume the same state — never replace QP cones or FOC.

Skip to “RL walks in sim” and you will discover that hardware’s first free walk is an unplanned impact experiment.

Failure gallery

What to freeze this week

Extend frames and your control config with:

• Who owns intended gait mode vs measured contact_set (and the disagree → abort/hold rule)
• GaitCommand / FootstepPlan fields + mid-rate tick budget
• How swing feet are tagged in WbcTaskSet every sample relative to contact_set
• Abort / shorten policy for stale, POWER_GATED, and low torque_scale
• Touchdown gate: ContactState thresholds aligned with Day 06 impact policy
• Twin hooks: contact delay, swing early-contact, sag under push-off — same as hardware
• Bring-up gate: no free walking / learned locomotion residuals until steps 1–5 above are green

The lesson is not “implement Raibert vs DCM vs a walking transformer.” It is make stepping the same contact-rich program in silicon and in the twin: one estimate, honest power ceilings, swing that does not fight support, and landings that inherit Day 07’s cones.

Next

Day 09 — Push recovery & disturbance rejection on the same EstimatorState + WBC ceilings: detect CoM / capture-point escapes, choose recover-in-place vs emergency step, and keep abort / soften paths when power or contact confidence collapses mid-recovery.