Training Coach
Doctrine
A report may cite any of these as doctrine-N.
doctrine-1Contract freeze and fingerprint disciplineThe policy I/O contract (observation layout, scales, history semantics, action pipeline) is frozen and fingerprinted; every exported policy is stamped and verified; contract changes ship as new versioned profiles that leave old artifacts bit-identical, and old policies run forever under their era's pinned profile.
Case. The 215-dim omni contract was frozen with a three-machine digest; the one contract-level extension (lateral feed-forward) went in as a new `omni_ff` profile with the old profile provably untouched, and the contract checker caught two real wiring bugs before any training (`contract-freeze-and-checker`). A silently changed gait-clock default would have fed old policies a 25% slower clock - closed by pinned legacy profiles (`legacy-profile-pinning`). A stale derived USD forked plant mass 2.2% until an automated source-vs-derived instrument gated it (`derived-asset-staleness-check`). A gain profile is part of the closed loop a policy was trained in and belongs in its stamp; the recovery line's anchored authority was left out of its manifest and recorded as the gap not to repeat (`gain-profile-belongs-in-the-stamp`), and a second policy behind a deploy-side switch made the handoff state itself a contract (`recovery-two-policies-and-a-state-machine`, `walk-recovery-fsm-handoff`).
Coach application. On any proposal touching obs/action semantics, defaults, or derived assets: demand the version/profile plan, the fingerprint update, and the checker extension in the same change; flag any old artifact that would run under new defaults.
doctrine-2Attribution by resolved training params - never eval-override knobsCapability differences between lineages are explained only by digging each lineage's *resolved* training configuration and eliminating columns; evaluation-side override knobs (kd-scale, power-scale, cycle-time) act on the plant for *every* policy and may serve as deployment mitigations but never as explanations.
Case. Low-friction robustness across 8 lineages x 3840 cells was traced to kd DR *bandwidth* - every lineage had ground friction pinned to (1.0,1.0), so "trained friction" could not be the axis; the parameter axis and the plant axis were explicitly separated after the first attribution conflated them (`kd-bandwidth-mu-law-attribution`). "Weak turning" on hardware was a power-scale plant effect, not a training gap (`deploy-knob-attribution-before-retraining`); slowing the deploy clock was out-of-distribution, not a feature (`cycle-time-override-is-ood`). The ground truth for what a run trained under is the logged per-run config, not the source tree (`resolved-config-is-source-of-truth`).
Coach application. Whenever asked "why is lineage A better", require the resolved-param table first; kill zero-variance columns; refuse explanations phrased in eval-knob terms; when a knob helps, label it deployment mitigation.
doctrine-3PASS gates become constraints; FAIL gates become objectivesOnce a skill passes its gate, that gate converts into a standing regression constraint (budget <= 2/20 against the parent baseline) for all later training; gates currently failing are the only legitimate objectives of the next rung.
Case. The C ladder ran one frozen 13-cell x 20-seed matrix at every rung with promotion = "new skill PASS and old skills within regression budget"; C1 was stopped and re-rooted precisely because it trained away the root's backward PASS (`fixed-acceptance-matrix-per-rung`, `preregistered-stop-criteria-per-rung`). The C4 product shipped only at 260/260 cells with zero regression.
Coach application. Keep the ledger: every PASS adds a constraint row; propose rungs only against FAIL rows; treat any constraint violation as stop-and-attribute, never "the next rung might win it back".
doctrine-4One variable per ladder rung - counted against what the checkpoint sawA rung changes one variable, where "one" is counted against the checkpoint's actual training state, not against the current config's diff; batching is allowed only when each change owns a disjoint symptom space with a pre-registered ablation order.
Case. Two rungs failed identically because resuming s1e-500 under the evolved config silently added four plant variables the checkpoint had never seen ("单变量纪律不只看「我改了什么」,还要看「checkpoint 见过什么」" - `resume-state-dr-audit`). v8 legally batched four orthogonal fixes with a written ablation order (`orthogonal-batch-with-ablation-order`); v9 spent one run completing a 2x2 factorial so either outcome convicted a factor (`fill-the-missing-factorial-cell`); v10b's three-way ablation wrongfully convicted the clock and had to be retried fairly.
Coach application. Before any resume: diff cfg against the checkpoint's logged training state. Before any batch: require the symptom-ownership map and ablation order in writing.
doctrine-5Pre-register risks, readings, and stop criteria before the ladderBefore a ladder or risky rung, write down the known risks, the interpretation of every plausible outcome, and hit-any-one stop criteria - frozen before training, tightened when priors say results should come fast.
Case. The C ladder opened with three numbered risks including the exact falsification condition for its own root choice; A/B arms carried "预注册读法(事后不改)" tables; a level expected to fail was run anyway for its pre-registered diagnostic value (`preregister-risks-and-fork-readings`). Stop criteria caught C4-redo rungs at +200 instead of full caps (`preregistered-stop-criteria-per-rung`); hardware sessions pre-registered per-config expected signatures and the disagreement rule "不改结论改账" (`preregistered-real-expectations`, `feasibility-accounts-lock-design-point`).
Coach application. Refuse to open a rung without the written risk/reading/ stop block; after results, read conclusions off the pre-registered table and flag any post-hoc reinterpretation.
doctrine-6Plant parameters are measured, never inventedEvery plant number carries measurement provenance: armature = N^2 x rotor inertia from no-load tests, friction split by rig and by API column, torque limits shaped by per-joint gait peaks, latency traced through the real pipeline, masses weighed - and DR bands are additive around the measured nominal, sized to the measured dispersion.
Case. Guessed friction was 2.5x low and guessed damping 5x high (`friction-measured-not-guessed`); armature had been 0 with a 9:1 gearbox (81x reflected inertia, `armature-n2-rotor-inertia`); a uniform torque derating was "the wrong shape" vs measured peaks (`torque-limit-shape-by-measured-peaks`); the delay implementation itself was a wrong plant for a whole lineage (`latency-lerp-reverse-extrapolation`); the run design point was locked by three accounts including the tau_limit/kd speed ceiling (`feasibility-accounts-lock-design-point`); identified friction had to land in the right simulator API columns to act at all (`sim-api-friction-columns`). The recovery and one-leg lines opened with the same kind of accounts before any reward existed - a connected static path and the torque along it for an armless get-up, and the gains single support needs to be holdable at all (`get-up-feasibility-accounts-before-training`, `single-support-gain-authority-probe`).
Coach application. For any plant value in a config review, ask "measured how?"; reject absolute ranges with no nominal; check API column mapping and derived-asset regeneration whenever measured values land.
doctrine-7Sim2sim gate before sim2real - under deployment conditionsEvery checkpoint passes a second, independently built simulator before hardware, and both the gate and the smoke loop run under the measured deployment conditions (real pipeline delay, honest contact parameters, the deployment gain/power profile).
Case. The standing order "先sim2sim 再sim2real" (`sim2sim-gate-before-sim2real`); acceptance flipped to match hardware only under measured condim/torsional friction (`eval-plant-honesty-contact-params`); gates moved permanently to `--delay 2` after the kicking incident (`pipeline-latency-is-plant-not-dr`); and the harness itself must be audited - a frame-convention bug in the cross-sim evaluator invalidated a whole line of verdicts (`body-frame-velocity-api-audit`). The recovery line's second simulator caught a torque penalty paid for by bracing the legs together (`torque-penalty-bought-by-leg-bracing`), and a 1.8x torque disagreement between the two plants stayed binding because its one surviving explanation was never tested (`torque-disagreement-between-simulators-unresolved`).
Coach application. Block any hardware request lacking a second-sim PASS at deployment conditions; when sim2sim and training-side metrics disagree, treat the evaluator as a suspect too.
doctrine-8Observation honesty - the actor's inputs are a hardware contractThe actor observes only signals the real robot produces with realistic noise; privileged truths go to the critic; history windows are estimators and must train under plant variation; rewards on quantities the actor cannot observe buy only average suppression, never closed-loop correction.
Case. Ground-truth velocity/forces went critic-only (`observation-honesty-critic-only`); frame_hist under zero DR memorized the trainer's plant fingerprint - 0/20 transfer (`history-obs-needs-plant-variation`); world-frame yaw rewards could not teach pull-back because heading is unobservable to the actor - correction was routed to the deploy outer loop instead of breaking the contract (`reward-observability-limit`, `deploy-heading-loop-and-align-training`).
Coach application. Audit every actor-obs element for hardware existence; require minimal plant jitter whenever history/recurrence exists; for each reward, ask "can the actor see this error?" and route correction tasks to outer loops.
doctrine-9Reward economics are audited in realized currencyReward design decisions are made on realized per-step magnitudes under the actual policy and command distribution: price the do-nothing optimum before adding a mode, compare achieved values to the computed ignore-floor, calibrate thresholds between measured healthy and sick distributions, and ship every new penalty with a withdrawal clause.
Case. feet_air_time at weight 2.0 realized 0.038 vs tracking 1.2 - drag was rational (`realized-contribution-audit`); ignoring a vy command cost 28-180x less than ignoring vx until a gated tracking term was added (`reward-cost-of-ignoring-audit`, `gate-new-reward-terms-by-command`); achieved-vs-floor separated "never learned" from "priced out" (`ignore-floor-diagnosis`); the foot-distance wall was placed between measured healthy (0.6% tax) and sick (55%) policies (`calibrate-threshold-between-healthy-and-sick`); the landing penalty carried a pre-registered stand-down condition and actually stood down (`calibration-threshold-with-withdrawal-clause`); two clearance terms were inert until zero-points and gate occupancy were checked (`inert-reward-term-audit`). A get-up policy sat because three gated terms paid the seated pose 84% of the return and the one term that could tell sitting from standing was an exp kernel reading 4.6e-5 at the real error (`seated-basin-dead-exp-kernel`); a torque-tail term was weighted by its measured steady value beside a peer term after the estimate proved 12x off (`tail-torque-needs-hinge-on-computed-demand`).
Coach application. Never discuss weights in the abstract: demand the realized-contribution table, the ignore-floor number, and the healthy-pay calibration before any reward edit is approved.
doctrine-10The zero-cost option must be the desired behaviorFor every penalty, name what the zero-cost option is; penalize failure events (slip, saturation excess, contact in flight windows), never the motion or joints that healthy behavior uses; make degenerate strategies fatal via termination where penalties cannot price them out.
Case. Joint-usage penalties for drift taxed a 1.4%-of-momentum channel 2.7/step and collapsed training; the slip penalty costs a non-slipping gait exactly zero (`penalize-the-slip-not-the-joint`). A frozen-at-clamp joint pays zero action-rate forever - only a pre-clip saturation penalty flips the cheat economics (`saturation-cheating-zero-rate-cost`). Ungated phase shaping made standing 42x more expensive than stepping and cooked the hip motors (`moving-gate-42x-stand-tax`); crouch-shuffling lived until a height termination deleted it (`termination-closes-degenerate-basin`). A gated penalty is an exit: the policy parked just outside an uprightness gate, then just under a height gate, to stop paying a stance tax, and only a positive band plus an always-on guard closed both (`penalty-gate-is-an-escape-hatch`); a soft-limit penalty that charged the standing pose itself bought a 4.1 deg lean (`soft-limit-penalty-charges-nominal-pose`); an unpriced foot attitude was spent on edge-standing (`unpriced-foot-attitude-is-a-free-variable`); and the one-leg line listed its cheapest cheats before training and still met one through a zero-gradient band (`enumerate-cheapest-cheats-before-training`, `binary-band-reward-fake-touchdown`).
Coach application. Run the "零代价的选项是什么" audit on every proposed term; convert motion taxes into event-conditional penalties; check the termination set against each known degenerate strategy.
doctrine-11Measurement discipline: independent referees, signs, distributionsA disputed measurement is adjudicated only by an independent algorithm from raw state; directional ability requires sign-antisymmetry under command reversal; bimodal metrics are reported as mode shares (never medians, never 3 seeds); ratios are not comparable when totals change; reward values compare only within one command distribution; single chaotic events never cross machines.
Case. The triple reversal - a good metric was "refuted" by a sibling metric that shared the disease (`independent-referee-for-metric-disputes`, `body-frame-velocity-api-audit`); same-signed +/- responses were bias, not turning (`same-sign-response-is-yaw-bias`); the swing median sat in a bimodal gap (`median-hides-bimodal-distribution`); "v6 is jitterier" died on absolute energies (`ratio-metrics-need-absolute-check`); yaw gain measured 15x wrong in an oscillating frame (`heading-integral-not-body-rate`); a 44% improvement evaporated under same-distribution comparison (`same-distribution-reward-comparison`); drift direction was a limit cycle (`multiseed-sign-test-for-drift`); a cross-machine push cliff was chaos (`single-impulse-recovery-is-chaotic`).
Coach application. Before accepting any surprising number: ask for the independent recomputation, the sign pair, the distribution shape, and the comparison conditions. Retract in writing when a metric falls.
doctrine-12The deployment pipeline is plantIrreducible pipeline properties - action latency, rate limits, power/torque scaling, teleop command mappings - are part of the nominal plant, modeled from day one and reproduced in every gate; deploy-side scalings are crutches that flag unmodeled plant, and they cannot be algebraically folded into training constants.
Case. Right-leg kicking was over-trained-delay x loop gain; power 0.8 was a gain-reduction crutch that retired when the delay was modeled (`pipeline-latency-is-plant-not-dr`); power derating damages non-forward axes first (`power-scale-hurts-nonforward-axes`); training at 0.4 scale as the "twin" of deploying 0.5 x 0.8 collapsed 0/20 (`deploy-scaling-not-training-equivalent`); one shared teleop speed sent an out-of-band lateral command and the robot clipped its own foot (`teleop-command-band-per-axis`); the latency DR range had not even covered the measured pipeline (`latency-dr-covers-measured-pipeline`). A rate limiter added at deployment only clipped a policy that kept commanding (`deploy-rate-limiter-windup`); moved into training and anchored on the last command it became an integrator in the balance loop (`slew-anchor-is-an-integrator`); anchored on the measured angle it bounded torque and kept the bandwidth (`beta-anchored-action-target`). The walking lines' safe setting, power-scale 0.8, cut the ends of the recovery policy's full-range travel and left its spikes alone; a gain inside the trained band did the job (`power-derating-cuts-full-range-contract`).
Coach application. Demand the measured pipeline latency/limits in the plant model and in gate conditions; treat every deploy-side derating as a question ("what is this compensating?"); block per-axis command sources that exceed training bands.
doctrine-13DR budget is finite; its distribution is the measured supportRobustness is a conserved budget: disturbance training on an already-hardened lineage borrows from existing margins; DR ranges span the measured deployment support - no fictitious tails (they buy degenerate gaits), no single constants (they allow thin-margin specialization); harden the plant only after the task distribution is final.
Case. The same push dose helped a narrow lineage and damaged a balanced one - budget conservation (`push-dr-conditional-budget-conservation`); wide latency tails bought drag-glide, constant values shipped 60% thinner tilt margins - the answer is a narrow band on the measured support (`dr-tail-plant-continuation`, `constant-value-dr-overfits-margin`); task-first ordering because hardening a soon-to-change task wastes budget (`task-shaping-before-plant-hardening`); COM randomization used deliberately as a behavior-shaping tool, and rolled back on symptom per its own contract (`com-randomization-forces-leg-spread`, `com-dr-rollback-on-symptom`). DR that is switched on can still be thin: the run policy fell in the frontal plane its gain-and-latency randomization never touched (`thin-dr-judged-by-channel-coverage`), and a friction priority settled under one action contract had to be re-measured under the next (`friction-priority-re-measured-after-plant-change`).
Coach application. Before any DR rung: check the untrained policy against the spec, the lineage's current DR load, and the measured real-world range; after it: audit retained margins, not just the new tolerance.
doctrine-14Gates measure what hardware feels: posture, margins, stripped assistsAcceptance batteries carry posture-class rows (tilt max median, per-joint L/R asymmetry, temperature) beside task rows, graded margin columns beside binary gates, chirality scored per side, at least one condition that removes the environment's free stabilization, and validated predictive scalars promoted into the gate.
Case. Three same-shaped judging errors - survival, displacement, wz-difference - all missed what the operator felt; posture metrics had the predictive power (`task-metrics-vs-posture-metrics`, `stand-gate-posture-not-survival`); binary survival saturated and hid a 60% margin gap (`constant-value-dr-overfits-margin`); v5 passed everything on the ground and failed suspended (`suspension-probe-removes-free-stabilizer`); the hip_roll (l+r) scalar predicted real drift direction and ordering and entered the battery (`hip-roll-sum-predicts-lateral-drift`); averages hide chirality (`chirality-scored-separately`); gait-quality gates are judged at speeds that demand a gait (`low-speed-commands-reward-dragging`). The recovery line added the rest of the kit: where failed episodes end, not only where they started (`end-state-confusion-matrix`); a frozen acceptance distribution with a pinned seed (`frozen-acceptance-distribution-and-pinned-seed`); video of the metric rollout itself (`video-as-acceptance-record`); and the admission that a 10 s episode cannot see a stance that fails after a minute (`episode-length-bounds-what-a-gate-sees`). The one-leg line removed a foot-spacing wall that no gate measured, and the feet met on hardware (`removed-wall-returns-on-hardware`).
Coach application. Review every battery for posture rows, margin columns, per-side scoring, and an assist-stripped condition; when operator feel and gates disagree, suspect the metric class first.
doctrine-15Fork and root selection: recoverability, maturity, frozen rewardsChoose fork roots by which candidate's deficits the coming training can pay back (precision is recoverable; lost plasticity, symmetry, and margins are not); prefer mature checkpoints as roots even when younger ones score better as products; never fine-tune through a reward change - continuation is legal only with the reward frozen and plant/DR widening one rung at a time.
Case. s1e-500 beat higher-precision candidates because its exclusive strengths were unrecoverable (`fork-root-recoverable-shortfall`); the b300 arm proved maturity is capital against adaptation shock (`root-maturity-vs-product-quality`); the B-arm scatter/half-recover/collapse signature falsified reward-change fine-tuning and drew the legal boundary for S2 continuation (`fine-tune-reward-change-falsified`).
Coach application. For root debates, build the exclusive-strengths table and ask "which side can be trained back?"; require dual-arm evidence for maturity claims; classify any proposed continuation as reward-frozen or not before approving.
doctrine-16Curricula: verified engagement, lineage counters, disease-phase gatingAutomatic curricula must prove they engage (a saturated ratchet is constant DR wearing a curriculum's name); every ramp counts lineage-cumulative progress, not per-process steps; penalties aimed at late-stage pathologies ramp in after exploration noise decays; difficulty rises on measured per-stratum success, never on schedule.
Case. The s1f ratchet capped at iter 248 and never engaged (`auto-curriculum-engagement-check`); the saturation ramp re-fired at +600 after every resume and no shipped product ever saw the penalty (`curriculum-counter-lineage-steps`); the same penalty worked once gated to the disease phase and became an untouchable mechanism (`gate-penalties-to-the-disease-phase`); record-high aggregate reward hid a fully-failing delay stratum (`aggregate-metrics-mask-subgroup-failure`); bucket share is not a gradient lever (`bucket-share-is-not-a-gradient-lever`). An assist curriculum keyed to a pooled success share was withdrawn on the strength of the categories that already worked (`curriculum-criterion-conditioned-on-lagging-category`); a pace set by per-step income moved only when that income was time-gated (`per-step-income-drives-speed-time-gate`), and the same gate had to be retired in a lineage without the disease (`time-gate-vs-wide-stance-retire-the-fix`).
Coach application. Ask every curriculum three questions: does it engage (show the internal state)? what does it count (process or lineage)? when is it present (against the pathology's phase)? Check where shipped checkpoints sit relative to every ramp.
doctrine-17Probe before training: feasibility first, hypotheses in tablesAfter two failed training attempts at a skill, stop training: demonstrate the behavior open-loop, enumerate hypotheses in a written table audited against actual configs cheapest-first, race one probe per side of the sim2real boundary for hardware-only pathologies, and use suspended tests to acquit or convict actuators before blaming authority.
Case. "在黑暗里试钥匙" - four sidewalk rungs failed until an open-loop probe separated exploration/waveform/authority in one experiment (`open-loop-probe-before-reward-tuning`); the foot-drag mystery fell to a seven-hypothesis config audit (`hypothesis-table-code-audit`); the period-doubling was resolved by racing a reward-side and a plant-side evidence line - and both paid off, one per sub-case (`period-doubling-evidence-race`); the suspended test acquitted the roll actuator in one measurement (`suspended-test-isolates-actuator-authority`). A read-only configuration probe told a wall from a slope in the recovery line's seated basin (`configuration-probe-wall-not-slope`), and the fix it pointed to - where the feet are - took prone from 0/159 to 158/159 (`prone-dead-end-is-foot-placement`); a knob that did not move its variable was recorded as no test of the idea (`dof-vel-penalty-is-not-a-pacing-knob`).
Coach application. When a skill resists training, prescribe the probe before any further reward edits; require verified target trajectories before imitation terms; keep a falsified-fixes list so closed roads stay closed (`amplitude-cut-falsified-yaw-fix`).
doctrine-18External advice is recomputed locally; values transfer as ratiosEvery external suggestion is classified adopt / already-have / modify / trap by recomputing its claim on the local reward table and probe data; numeric values transfer only as dimensionless ratios (to tracking weight, leg length, sqrt(gL), control rate); citations are verified to exist.
Case. "Start vy very small" would have destroyed sidewalk learning on this reward table - the gradient scales quadratically (`external-advice-audit-against-own-arithmetic`); swing-height targets and weights transferred correctly only through leg-length and tracking-ratio scaling (`transfer-ratios-not-absolutes`); the "6-step delay" was refused for lacking a control rate (`latency-dr-covers-measured-pipeline`); a borrowed reference's structure was FK-verified and its amplitude re-derived from the division of labor (`reference-structure-fk-amplitude-division`); retrieval agents fabricated verbatim arXiv quotes - only source-verifiable material was used; and one dismissed suggestion later proved right for a different mechanism, and was credited (`cycle-average-tracking-for-gait-quantities`). An advisor's staged state machine turned out to exist in none of the three papers it cited, and reading them changed the plan (`advisor-paraphrase-vs-paper`).
Coach application. Intercept every "paper X does Y" with the local recomputation; convert absolutes to ratios before comparison; verify quotes; revisit dismissed advice when new mechanisms appear.
doctrine-19Hardware sessions are scripted experiments, not tuning sessionsReal-robot time executes a pre-registered matrix: risk-ordered (baseline first, fragile last with a spotter), stage-gated (suspended smoke before ground), A/B sessions bracketed by a repeated reference run, operators briefed on measured zero-command and untrained-axis behavior, chirality-aware disturbance protocols, no field tuning - the only legal field changes are scripted, single-variable, and self-reversing.
Case. The S2 acceptance sheet (`risk-ordered-real-deployment`, `battery-bracketed-real-ab`, `know-zero-command-behavior`, `push-test-chirality-protocol`, `no-field-tuning-protocol`); the RAM-only torque experiment with automatic power-cycle rollback (`reversible-single-variable-field-experiments`); and the sim-veto rule - even sim's condemnations get one safeguarded hardware check when they judge the purpose-built configuration (`sim-veto-needs-real-confirmation`). The recovery line's first real run went ahead with its preconditions unmet and was stopped as dangerous (`first-real-get-up-violent-stage-one-policy`); after it: a staged hang, mat and floor protocol (`staged-hang-mat-floor-for-get-up`), a fixed power-cycle pre-flight and two-machine discipline (`power-cycle-preflight`, `two-machine-config-discipline`), a fall guard replaced rather than switched off (`fall-guard-becomes-a-state`), and logs that are part of the run (`hardware-log-is-the-attribution-input`).
Coach application. Turn every hardware request into a runbook with order, gates, brackets, briefing, and anomaly plays; refuse improvised parameter changes on the floor.
doctrine-20Close questions in writing; restart when the debt is structuralAudited questions get frozen verdicts with citable wording and an explicit reopening bar; hardware verdicts are dated by deployment-stack and calibration state and expire when those change; and when successive rungs shuffle symptoms without net progress, freeze the lineage as regression baselines, pay the structural debts, and retrain minimal - carrying laws and instruments, not weights.
Case. The chirality and COM questions were closed with frozen wording and "no reopening without new hard evidence" (`frozen-verdicts-semantic-boundaries`); v5/v6's condemnations expired with the deploy stack (`stale-verdicts-under-old-stack`); a 2-degree calibration fix moved the whole runnable envelope (`zero-offset-calibration-shifts-envelope`); plant upgrades are era boundaries with paired re-baselining (`plant-swap-invariants-vs-shifts`); and the 2026-08-05 reset froze v5-v11, fixed the latency FIFO / manifest / sampling / reward-table debts, and restarted - producing the lineage that reached hardware SOTA (`freeze-lineage-fix-structure-restart`, `minimal-reward-table-with-provenance`). The recovery line's real-robot verdicts ended up in three places that disagree, one of them an undated note in a command file (`write-hardware-verdicts-back`).
Coach application. Maintain the closed-questions ledger and quote it when symptoms recur; stamp verdicts with stack/calibration versions; when a team is three rungs into symptom-shuffling, raise the restart question explicitly with the freeze-fix-restart pattern.
doctrine-21Name the quantity in the space it lives inA goal, reward term or acceptance criterion about the feet, the base or the contact state is computed from the quantity itself - world poses, forces, per-category outcomes - never through a joint-angle, single-signal or pooled stand-in that assumes everything else sits at nominal; and every detector is validated on a behaviour known not to contain the event before it becomes a gate.
Case. The recovery line was caught three times: |ankle roll| as "flat feet" sold stance width and the real robot slid into the splits, a hip-roll criterion was confounded by 50 deg of yaw, and the joint table said 0.271 m where the feet were 0.159 m apart; task-space terms produced the first flat, wide stance (`joint-space-proxy-for-task-space-quantity`). Flight detection lied in both directions across two lines - foot height flagged 40% false flight on a walking gait, contact force alone flagged slip chatter as hops (`contact-detector-single-signal-lies`). A pooled height average described a robot that did not exist - six in ten standing, four in ten sitting (`zero-partial-credit-is-not-an-iteration-problem`) - and the walking line had learned the same lesson on yaw rate (`heading-integral-not-body-rate`).
Coach application. For every reward term and gate row, ask what physical quantity it stands for and whether it is measured directly; flag joint-space or single-signal stand-ins for task-space goals, ask for a detector validated on a negative control, and split pooled metrics by category before reading them.
doctrine-22Continuation needs a live gradient; a release is chosen by a scanContinue a converged policy only on a change that creates a live gradient, on a short budget, with every checkpoint scanned on the transfer axis; choose a release by running the full battery over a band of checkpoints and stop on signals, never by taking the last one; and when edits to the terminal phase cannot move a behaviour, roll back and retrain with the constraint present from the start, keeping the order in which the lineage acquired its mechanisms as explicit curriculum phases.
Case. A continuation with no new gradient drifted MuJoCo transfer from 100/98% to 80/28% while every Isaac gate stayed perfect, and a live-gradient continuation at the same depth kept it (`converged-continuation-is-poison`). One-leg checkpoints 100 iterations apart failed 1 and 38 of 40 cells, and late ones degraded (`checkpoint-choice-is-a-full-gate-scan`). Four in-lineage stance fixes failed because the stance was the end of the get-up path, and from scratch it grew right (`stance-decided-by-get-up-path`); fixes stacked on degraded states were rolled back by the user (`stop-stacking-roll-back-and-audit`); and the lineage's final recipe, trained from scratch in one run, sat at 0% because the order of its curriculum was part of the product (`curriculum-history-is-part-of-the-product`). The omni line's short adaptation budgets and mature roots are the same law seen from the other side (`continuation-budget-not-from-zero`, `root-maturity-vs-product-quality`).
Coach application. Before approving a continuation, ask for the new gradient, the budget and the transfer axis in the scan; before approving a release, ask for the scan; after three rungs without progress on the target, propose rolling back to the last good checkpoint and a from-scratch phase plan instead of a fourth patch.
Experience cards
125 cards matching “deploy-scaling-not-training-equivalent”.
Training at scale 0.4 is NOT the twin of deploying 0.5 at power 0.8 - the algebra matches, the learned policy does not
deploy-scaling-not-training-equivalentNever assume deploy-side scalings can be folded into training-time constants ("burning the crutch into training"): the learned optimum depends on the training-time authority, so treat such conversions as full experiments with pre-registered expectations and a sim2sim gate before any hardware.
Symptom
s1g (S1.6) trained from zero at action_scale 0.4 - meant as the "training twin" of the hardware-proven s1c-at-power-0.8 (0.8 x 0.5 = 0.4) - was all green in Isaac (zero falls, reward 117) yet scored 0/3 across all eight checkpoints and 0/20 at 20 seeds in the MuJoCo gate, falling forward at median 1.57 s with a 2.9x speed overshoot.
Context
The pre-registered expectation (survival gate should pass, since the conviction matrix showed s1c@0.8+delay2 all-survive) was cleanly falsified, and the harness was acquitted by controls: --delay 0 fell identically (not a delay fragility), check_contract all green, and s1c through the same harness survived 2/3. The verdict: "「s1c@0.8 = 0.4 训练孪生」的代数等价不成立" - a policy deployed with a derated output still LIVES in the 0.5 internal model it trained under (its value function, its expectations of its own authority), while a policy that starts training with reduced authority learns a different, clip-hugging gait with zero margin for plant differences ("部署端打折的策略活在 0.5 的内模里,训练起点收权限学出的是贴 clip 的 另一套步态,对 plant 差异零余量"). Result: the policy was withdrawn before hardware ("撤回——不上真机"), the lineage root moved back to the 0.5-contract s1c-5500, and this became the C ladder's cited fact-check ("s1g 是 0/20 证伪出局的那一代").
Change
The amplitude-surgery route abandoned; contract kept at scale 0.5; the deploy-side 0.8 crutch later retired on its own merits when the delay-complete s1e generation ran at full power.
Outcome
One training run bought a clean falsification of a plausible algebraic identity; no hardware time was spent on it because the sim2sim gate caught it.
Mechanism
Output scaling commutes with the network arithmetic but not with learning: the training-time scale shapes which gait solutions are reachable and how much clip headroom the optimum keeps. A derated mature policy retains the wide-authority solution executed softly; a from-zero narrow-authority policy finds a different optimum that saturates its smaller envelope - the two are not the same controller in different units.
Applies when
- proposing to move a deployment derating into a training constant
- a scaled-down contract policy hugs the action clip
- Isaac-green / cross-sim-zero results on a re-scaled lineage
“预注册 a) 证伪——Isaac 全绿(零摔/reward 117)但 MuJoCo --delay 2 八档 checkpoint 扫描全数 0/3、iter6500 20-seed 0/20 … 「s1c@0.8 = 0.4 训练孪生」的代数等价不成立: 部署端打折的策略活在 0.5 的内模里,训练起点收权限学出的是贴 clip 的另一套步态,对 plant 差异零余量。”
train/OMNI_V0_SPEC.md § 3. S1.6 判决(2026-08-07 验收) An outer heading P-loop at deploy cut drift 10x because its output stays inside the trained command band - then training was aligned to it
deploy-heading-loop-and-align-trainingFix drift-class problems first with an outer loop whose output provably stays inside the trained command band; when adopting it permanently, align the training command generator to the deployment's actual command mixture (feedback-driven AND constant), matching law, gain, and clip exactly.
Symptom
Persistent heading drift on straight-line walking (v6 net yaw 60.3 deg over 15 s) that reward-side fixes had only partially tamed.
Context
The deploy stack added --heading: an external P loop wz = clip(0.5 * wrap_to_pi(theta0 - theta), +/-0.6), recomputed each frame and fed into the policy's ordinary wz command slot. Measured: net yaw walk_v6 60.3 -> 5.8 deg, walk_v5 17.4 -> 4.2 deg. A run-level audit later corrected the mechanism story: training had heading_command=False since v1 - the policy had NEVER seen heading-error feedback, so the loop works purely because its output lands inside the trained command distribution wz ~ U(+/-0.6): "收益真实,当时的机理解释写错了" (the benefit is real; the mechanism explanation had been wrong). v8 then closed the loop properly: training-side heading command enabled with rel_heading_envs=0.5 - half the envs get heading-error-driven wz, half get explicit constant wz, because deployment feeds wz BOTH ways (straight-line = heading feedback, turning = constant command) and rel=1.0 would have made constant-wz turning out-of-distribution. The law, gain, and clip were aligned item-by-item between trainer and deploy tool.
Change
Deploy-side outer loop first (no retrain needed); then v8-D enabled the matching training-side heading command at rel=0.5 with identical gain (0.5) and clip (+/-0.6), contract unchanged (wz slot carries the computed value).
Outcome
Drift handled at deploy (5.8 deg) generations before training caught up; the alignment removed the residual train/deploy distribution mismatch, with the accepted cost booked (open-loop straight walking becomes more OOD for heading-envs - irrelevant since acceptance and deployment always run the loop).
Mechanism
A learned velocity-tracking policy is a valid inner loop for any outer controller whose commands stay within the trained command distribution - the policy needs no knowledge of the outer objective. Full alignment then requires training on the same mixture of command sources the deployment actually uses, in the observed proportions.
Applies when
- heading/position drift on a velocity-tracking policy
- designing outer loops over learned locomotion controllers
- training command distribution differs from how deployment feeds commands
“审计更正(2026-08-02,run 级 env.yaml):训练侧自 v1 复盘起就是 heading_command=False … 策略从未见过航向误差反馈。--heading 是评估/部署侧外加的航向 P 环(wz=clip(0.5·err,±0.6), 落在训练分布 wz~U(±0.6) 内)。实测净偏航 walk_v6 60.3° → 5.8° … 收益真实,当时的机理解释写错了”
train/WALK_V7_SPEC.md § 0. 本轮之前已经改掉 (航向闭环, 含审计更正) Deployment power derating damages non-forward axes far more than forward - sweep it in sim before deploying
power-scale-hurts-nonforward-axesTreat deployment power/torque scaling as a plant parameter: evaluate the policy in sim at the exact deployment scale, expect non-dominant axes to degrade first under derating, and either deploy at the training power or train with power randomization.
Symptom
Policies deployed at power-scale 0.8 (a safety derating of commanded torque) looked fine walking forward but were weak at backward and turning, inviting the wrong diagnosis "the skill was not trained well".
Context
Measured repeatedly: on s1e, going 1.0 -> 0.8 cost forward 18% but backward 58%; on C4-ff800, turn tracking was +25%/+40% at pw0.8 vs +75%/+58% at pw1.0, backward 51-52% vs 97-103%, while forward stayed 96-98% at both. Sim evaluation numbers in the plan were all pw1.0, but the robot was being run at 0.8.
Change
Pre-deploy protocol added: sweep the exported policy across power in sim (for PW in 0.8 0.9 1.0: eval_c_matrix --power $PW --seeds 20) and deploy at the first level where both turn directions reach >=50%. For C4 the recommendation was raise the robot to pw1.0 - the sweep showed it nearly free (saturation 47%->33%, left foot-clipping danger zone 25%->6%, cost only tilt 6.7->8.3 deg).
Outcome
Turning "weakness" resolved without any retraining; the sim sweep correctly predicted the real-robot signature at both power levels.
Mechanism
Forward walking is the reward-dominant, torque-cheapest skill with the most margin; backward/turn/sidewalk live closer to the torque envelope, so a uniform torque derating consumes their margin first. Training ran at power 1.0 (the trainer does no power scaling), so deploying at 0.8 is a systematic underactuation the policy never experienced.
Applies when
- deploying with any torque/power derating or safety scale
- secondary skills (backward, turn, lateral) underperform on hardware while forward walking looks fine
- choosing the deployment power level for a new policy
“power 衰减对非前进轴的伤害远大于前进轴（s1e：前进 1.0→0.8 掉 18%，后退掉 58%）。转向是非前进轴，0.8 下很可能明显跟不动。”
train/C_LADDER_RUN.md § 3c. A-2 上机前先定部署力度档 / 3p. 二 Slowing the gait clock at deployment is out-of-distribution and backfires - lower the commanded speed instead, or train the knob
cycle-time-override-is-oodAny deployment override must correspond to a dimension the policy was trained to handle; to make a parameter field-adjustable, randomize it in training and observe it - otherwise use the levers inside the trained envelope (commands) and leave the knob alone.
Symptom
Real-robot feedback "walks very fast and unstable" suggested slowing the gait; a deploy-side --cycle-time override existed, making "just slow the clock" a one-flag temptation.
Context
A sim sweep of the override on walk_v6 @cmd 0.3 showed monotone degradation away from the trained 0.40 s cycle: at 0.50 s tilt jumped 7.9 -> 13.2 deg and landing force 1.52x -> 2.24x; at 0.80 s (half speed) clearance collapsed to 3 mm - dragging again - with 20 deg tilt. Meanwhile the legitimate lever, lowering the commanded speed with the clock untouched, improved everything monotonically: cmd 0.1 gave 104% tracking, 6.7 deg tilt, minimum slip - the most stable operating point. The file distinguishes the two "slows" explicitly: lower command = smaller steps at the same 2.5 Hz rhythm; a slower rhythm itself requires retraining - randomize cycle_time (e.g. 0.40-0.65 s) during training and expose it as an observation, and only then does --cycle-time become a field-adjustable knob.
Change
Deployment guidance: never ship a cycle-time override the policy was not trained under; respond to "too fast/unstable" with lower commands; schedule clock variability as a training-time (contract-level) change if a field knob is wanted.
Outcome
The sweep quantified the trap before hardware paid for it (dragging and 2.2x landing force at slowed clocks); cmd 0.1 documented as the stable demo point.
Mechanism
The policy is a function fitted around the training distribution; a deploy-side override moves an input (phase rate) to values never seen, so behavior degrades unpredictably - the knob LOOKS like a capability because it exists in the code, but capability lives in the training distribution, not the interface.
Applies when
- a deploy tool exposes overrides (clock, scale, gains) beyond the training distribution
- hardware feels "too fast/aggressive" and a quick knob exists
- deciding between a deploy-side tweak and a retrain
“0.80s | 1.25Hz | 0.165 | 3mm(拖地) | 20.0° … 慢一半直接崩 … 策略按 0.40 训练, 别的周期属分布外。… 降指令速度才是有效杠杆 … cmd 0.1 是最稳的工作点。… 要节奏本身变慢必须重训 —— 训练期把 cycle_time 随机化(如 0.40~0.65s)并作为观测的一维, 部署时 --cycle-time 就成了现场可调的旋钮。”
train/WALK_DIAGNOSIS.md § 2026-08-01 追加: 调慢步态时钟(--cycle-time)在仿真里是反效果 The deploy-side walk/recovery switch - into recovery at tilt > 65 deg held 0.3 s, back at tilt < 15 deg with angular rate < 1 rad/s and straight knees held 1 s, a 15 s timeout, last action cleared both ways - and the handoff steps the design required are only partly implemented
walk-recovery-fsm-handoffSpecify a deploy-time controller switch as hysteretic, time-filtered predicates the robot can measure (proxy what it cannot, e.g. straight knees for height), a timeout that ends in a safe stop, and a complete handoff (history, clock, last action, command ramp) - then test that the code performs every handoff step, because the design document is not the implementation.
Symptom
With a recovery policy and a locomotion policy as separate networks, the robot needs a switch: when is it "fallen", when is it "up", and what state must be reset so the next policy does not act on the previous one's history.
Context
The 08-09 design: enter recovery when fallen (tilt > 55 deg or height < 0.60 x 0.384 m) for 150 ms, leave for a stand-hold when upright (tilt < 12 deg, height > 0.85 x 0.384 m, feet steady, |omega| < 0.8) for 400 ms - wide entry, strict exit, hysteresis - then a mandatory handoff trio before walking resumes (reset the walking policy's observation history, restart its phase clock at 0, clear its previous action and latency buffer) and a command ramp instead of a jump. The 08-14 implementation in deploy_policy (--recovery-policy): each policy under its own manifest contract (walk: nominal + scale; recovery: beta-anchored), the same rl_default gains, the tilt cutoff disabled; RECOVERY at tilt > 65 deg for 0.3 s; LOCO again at tilt < 15 deg and |omega| < 1 and knees straight (< 0.35 rad) for 1 s - the robot's computer has no height estimate, and straight knees stand in for height so a V3.0-style upright kneel cannot pass as standing; RECOVERY longer than 15 s ends in a safe stop; last_action cleared on both switches; command forced to 0 during RECOVERY; power scaling applies to LOCO only.
Change
An open account was written down with it: the recovery end state (0.355 m stance, hip yaw -/+27 deg) is outside the walking policy's training start distribution, so the first test must use stand / zero command as LOCO, and the long-term fix is to widen the walking policy's initial states rather than bend recovery's stance to suit walking.
Outcome
The spec records only a successful compile (py_compile); the hang test was left to be done on site. The operator runbook carries the three-step procedure (hang with stand as LOCO, mat and push, then omni walk as LOCO) but no outcome.
Mechanism
Two policies trained separately each assume their own history, clock and last action; a switch that carries any of them across feeds the next policy a state it never saw - the design called this "the walking policy seeing a ghost history".
Conflicts
The 08-09 design requires resetting history, clock and previous action plus a command ramp; the 08-14 implementation records last_action clearing and a zero command during recovery; the one-leg spec of 2026-09-14 lists the handoff hygiene as specified but not implemented - reset_history() is called by nothing (a 215-dim policy would carry four frames of pre-fall history), the phase clock is not zeroed, and there is no command ramp back to LOCO. No hardware run of the FSM is recorded in either source.
Applies when
- switching between separately trained policies on hardware
- a policy with history or phase observations is re-enabled mid-run
- the robot lacks a sensor the switching criterion was designed around
“**判据**:进 RECOVERY = 倾角 >65°(`--fall-tilt-deg`)持续 0.3 s;回 LOCO = §5 真机可测子集:倾角 <15° ∧ |ω|<1 ∧ **膝直 <0.35 rad(NX 无高度观测, 高度门用膝直代理 —— 防 V3.0 型"跪坐但直立"误判)** 持续 1 s (`--recover-hold`);RECOVERY 单次 >15 s(`--recovery-max-time`)安全停。 … **切换卫生**:两向切换 last_action 清零;RECOVERY 态 cmd 强制 0”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §50 FSM 双策略调度(2026-08-14,用户令):deploy_policy --recovery-policy Set torque limits per joint from measured gait peaks - a uniform percentage is the wrong shape, and training must use the deployed numbers
torque-limit-shape-by-measured-peaksMeasure per-joint torque peaks in the actual gait and set each limit as measured-peak x margin capped at rating; then propagate the same numbers into training and add an automated deploy-time consistency check - never derate by a uniform percentage, never let training assume torque deployment will not grant.
Symptom
A uniform 50% torque derating (18/8.5/7) had piled safety margin on the joints that never use it while cutting the busiest joint below half its measured demand.
Context
Per-joint gait peaks were measured (walk_v5 at cmd 0.3/0.6): RS06 (hip_pitch/knee) uses 5.5-5.9 N*m = 15-16% of its 36 N*m rating - cutting it to 12 is a free safety win; RS02's ankle_pitch runs at 16.2 N*m = 95% of its 17 N*m rating - "它是速度的硬件瓶颈", no room to cut; RS00 measured 36-44%, capped at 11. The resulting shape 12/17/11 replaced the uniform percentage. Sweeps across several limit sets (rated / 50% / 14-17-11 / 12-17-11) produced identical speed, lift, and landing force - within this range the limits do not shape the gait; what matters is consistency: "关键是训练和硬件必须是同一个数", because the exporter fills effort_limit from tau_limit, and a policy trained at rated 36/17/14 "会假设有三倍力矩可用" while deployed at 12/17/11 (exactly the v5 cross-generation inconsistency later suspected in its wild kicking).
Change
robot.yaml tau_limit set to the measured-shape 12/17/11, firmware written to match, and train/isaac_values.py regenerated so training sees the same limits; the deploy tool self-checks limits against robot.yaml on every run.
Outcome
Free safety margin captured where demand is low, the real bottleneck joint left at rating, and the train/deploy torque worlds unified with an automated consistency check.
Mechanism
Torque demand is grossly unequal across joints in a gait (15% vs 95% of rating here); a uniform percentage misallocates the safety budget by construction. And since the trainer treats effort_limit as a plant truth, any train/deploy mismatch is an invisible plant gap of exactly the mismatch ratio.
Applies when
- choosing safety torque limits for a legged platform
- training-vs-deployment actuator limit audit
- one joint runs near rating while others idle
“曾用统一 50%(18/8.5/7)是错的形状: 把余量堆在用不到的 RS06 上, 却把 ankle_pitch 砍到需求的 52%。… RS02 在 0.6 m/s 已用到额定 95%, 它是速度的硬件瓶颈 … 实测多组限幅 … 完全一致 —— 限幅在这个范围对步态零影响, 关键是训练和硬件必须是同一个数。… 若训练仍按额定 36/17/14, 学出的策略会假设有三倍力矩可用。”
train/WALK_V6_MINIMAL.md § 3. 训练侧必须同步的一件事 Tightening the bridge's rate limiter under an unchanged policy cut torque peaks 30-50% and made other things worse - the policy cannot see the limiter, keeps commanding and winds up; a deploy-side limiter is a safety net, not a cure
deploy-rate-limiter-windupA rate or torque limiter added at deployment lowers peaks but the policy still commands as if unconstrained (saturation, windup, new contacts); use it as a safety net mirrored in evaluation, and put the constraint where the policy can learn around it.
Symptom
After the violent first real-robot get-up, the cheapest candidate fix was to tighten the bridge's slew (rate) limit for the recovery policy without retraining.
Context
Probe on R3.1 in MuJoCo (5 categories x 3 seeds, mu 1.0), monkeypatching the limiter with no repository change: TIGHT = RS06 4.0 / RS02 3.0 / RS00 2.0 rad/s (about 0.08/0.06/0.04 rad per policy step) against the current vel_limit setting.
Change
The probe decided the role of the limiter rather than a deployment.
Outcome
Success 14/15 -> 12/15; get-up median 2.35 -> 3.53 s (max 9.30); torque demand peak median hip_pitch 164% -> 111%, knee 166% -> 86%; action saturation still 100%; leg-leg contact 558 -> 860 frames. The limiter was kept only as a real-robot safety net (mirrored into sim2sim evaluation); the cure moved into training - where the next lesson was that a limiter anchored on the last command is itself an integrator (slew-anchor-is-an-integrator).
Mechanism
A policy that never trained with the limiter keeps issuing the targets it learned; the limiter clips them, the target window runs ahead (windup), and the robot follows a trajectory the policy never evaluated.
Applies when
- a trained policy is too violent on hardware and a quick deploy-side fix is tempting
- adding slew, torque or velocity limits in a bridge or firmware
- evaluation and deployment use different limiter settings
“判读:**链路侧收紧立等可取地把 τ 峰值砍 30~50%,但成功率掉、饱和率仍 100%、 腿-腿接触反升** —— 策略感知不到限速器,目标窗口继续狂奔。⇒ 收紧 slew 只配当 **真机侧安全网**(必须同步进 sim2sim 口径,基础设施现成),**不配当治法; 治法必须进训练**。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §26 探针:收紧桥层 slew,r3_1 不重训直接测 Late training leaned on sampling noise as a stability crutch - deterministic play collapsed while training metrics stayed green
noise-crutch-deterministic-collapseEvaluate the deterministic policy in an external harness on a fixed cadence during training (not just at the end), select checkpoints on that curve, and treat a collapsing noise_std with rising training reward as a warning that noise is load-bearing.
Symptom
omni_s1's final checkpoint (model_5999) fell at 4 s even in Isaac's OWN deterministic play, while checkpoints from iter 1700-4000 were fine - and no training metric flagged anything. Policy noise_std had collapsed to 0.045 by iter ~990 (final 0.033).
Context
Diagnosis: the policy had learned to use its exploration noise as a dither/stabilizer - "策略把采样噪声当稳定拐杆,训练指标看不见" (the training metrics cannot see it, because training always runs with noise on). Countermeasures: entropy_coef 0.005 -> 0.01 to slow the std collapse, and - the structural fix - an in-training smoke loop (watch_ckpt.py): every 500 iters, export ONNX directly, run 3-seed MuJoCo evaluation, log CSV/TensorBoard curves plus three-view videos. The doctrine line was written in bold: "训练指标全绿不再是发育健康的 证据,冒烟曲线才是" - green training metrics are no longer evidence of healthy development; the smoke curve is. The follow-up run s1b showed the drift metric follow a U-shape (73 -> 8.6 at iter 3500 -> 76), making checkpoint selection BY the smoke curve (early stop at 3500) the shipping mechanism, with terminal re-degradation booked as known and unresolved.
Change
entropy floor raised; watch_ckpt smoke loop instituted as standing infrastructure; checkpoint selection moved from "last iteration" to "best point on the deterministic smoke curve".
Outcome
s1b shipped from iter 3500 (the U-bottom) instead of a degraded terminus; every later lineage (s1c/s1e, the C ladder's --every 100 loops) inherited the watcher as the standard guardrail.
Mechanism
PPO evaluates and improves the stochastic policy; if noise itself stabilizes the gait (dither smoothing a marginal limit cycle), the deterministic mean policy is a different, worse controller that training never measures. External deterministic evaluation on an independent simulator is the only readout of what will actually be deployed.
Applies when
- final checkpoints underperform mid-training ones
- noise_std collapses early while training reward climbs
- deciding which checkpoint to export and ship
“训练后期确定性脆化——noise_std iter~990 收到 0.045(终 0.033),model_5999 连 Isaac 确定性 play 都 4 s 摔(1700~4000 正常):策略把采样噪声当稳定拐杖,训练指标看不见。对策:entropy_coef 0.005→0.01 + train/watch_ckpt.py 训练中冒烟曲线 … 训练指标全绿不再是发育健康的证据,冒烟曲线才是。”
train/OMNI_V0_SPEC.md § 3. S1.1 修订记录 ② When hardware underperforms, audit deployment knobs before prescribing retraining
deploy-knob-attribution-before-retrainingBefore any "retrain it" decision, reproduce the symptom in sim under the exact deployment configuration; if the symptom follows the deployment knob rather than the checkpoint, fix the knob or randomize it in training - never top-up-train the skill.
Symptom
Real-robot feedback after the C4 deployment - "turning is weak" - with two retraining options on the table: top up turn training, or restart from the s1e root.
Context
The sim account showed the policy turned well (75-81% at pw1.0); the robot was deployed at power-scale 0.8. The 3-6 pp difference between C2 and C4 policies at the same power was noise; the 40-50 pp difference between power levels was the entire effect. Both proposed retraining paths would have burned budget on a non-existent training gap, and restarting from s1e would additionally have discarded the sidewalk skill that took four rungs and a coordinate-bug hunt to obtain.
Change
Decision: retrain nothing. (1) Try pw1.0 on hardware first - sim says net gain; (2) only if 1.0 is unacceptable (heat/feel), the correct training fix is power/torque randomization in the S2 plant line (one variable, fixes turn and backward together) - not skill top-up; (3) restart-from-root explicitly ranked worst.
Outcome
The "weakness" was fully explained by the deployment knob; the sim/real signatures matched the earlier power-derating law verbatim ("与 C2 时代 power 衰减主要伤非前进轴 逐字吻合").
Mechanism
The policy's competence is defined under its training plant; deployment knobs (power scale, teleop mapping, command bands) silently define a different plant. Attributing a deploy-plant effect to a training gap produces exactly the wrong fix - more training on the wrong variable.
Applies when
- real robot underperforms a skill that sim says is fine
- proposals on the table include retraining or re-rooting
- deployment uses any override the trainer never saw (power scale, remapped commands, different control rate)
“正确的训练修法不是补训转向，而是训练时加 power/力矩随机化让策略在 0.8 下自己补偿 —— 单变量，属 S2 plant 线，一次同时修好转向与后退；从 s1e 重训是最差选项：丢掉四轮 + 一个指标 bug 才换来的侧走，而 C2 的转向本来就没问题。”
train/C_LADDER_RUN.md § 3p. 三 处置顺序（回答「补训转向 还是 回 s1e 重训」：都不该） When training fails repeatedly, inject the target behavior open-loop - stop tuning rewards for an unverified behavior
open-loop-probe-before-reward-tuningAfter two failed training attempts at a skill, stop training: demonstrate the behavior open-loop on the real plant/sim first, and only resume training once you hold a measured, safe, sign-verified target trajectory.
Symptom
Three sidewalk training rounds failed; hypotheses multiplied (exploration failure / wrong reference waveform / insufficient authority) with no way to pick between them by running more training.
Context
Instead of a fourth reward guess, the team wrote probe_side_ref.py: the candidate reference is injected open-loop on top of a frozen policy's output (bypassing PPO entirely), directly measuring "what happens if the robot literally does this waveform" - separating all three hypotheses in one experiment (5 seeds x 8 s per condition, several waveform families and gains). The probe immediately eliminated the authority hypothesis (full-amplitude execution, 5/5 survival) and localized the problem to the waveform/measurement side. The closing principle was written down after the saga: without a verified target behavior, tuning rewards is "在黑暗里试钥匙" (trying keys in the dark).
Change
Standing method: before opening another training rung on a failing skill, build an open-loop (or task-space) generator of the intended behavior, measure whether the physical system can express it and what it looks like - then train toward a verified, quantified target.
Outcome
The probe chain produced the verified waveform (reversed-sign triangle, half gain), quantified safe amplitude (tilt 8.2 deg at band top, foot distance clear of the wall), exposed the metric bug when probe and training disagreed, and ultimately supplied the feed-forward that made C4 pass in +100 iters.
Mechanism
Training couples exploration, reward design, and feasibility into one opaque outcome; open-loop injection cuts the loop and tests feasibility and waveform alone. A behavior demonstrated open-loop converts the remaining failure into a pure credit-assignment/reward question - and its measured trajectory becomes the reference itself.
Applies when
- repeated training failures on one skill with multiple live hypotheses
- uncertainty whether the platform can physically express the behavior
- a reference trajectory's shape/sign/amplitude is guessed, not measured
“三轮 FAIL 之后不再猜，写 train/probe_side_ref.py 把参考开环注入到策略输出之上（绕过 PPO），直接量「照这个波形做会怎样」，一次分开三个假说：甲 探索 / 乙 波形 / 丙 权限。”
train/C_LADDER_RUN.md § 3f. C4 真因定谳（开环探针） / 3k. 建议的下一步 Add a single-point-suspension test to acceptance - the ground is a free stabilizer that hides divergence
suspension-probe-removes-free-stabilizerInclude at least one acceptance condition that strips the environment's free stabilization (suspension, or equivalent) - the sim-passing policy that fails on hardware is often failing a condition the battery never posed.
Symptom
walk_v5 looked healthy in every on-ground sim test yet diverged on the real robot - the acceptance battery had never measured a condition that would have revealed it.
Context
The battery gained a single-point-suspension probe (robot hung, feet free): measure torso tilt while the policy runs without ground contact. v5 scored 45.9 deg mean tilt suspended - wildly unstable - which the file calls "最灵敏的失稳探针(拿掉地面这个免费稳定器)": ground reaction forces passively stabilize a marginal policy, so on-ground metrics saturate long before the policy's internal balance is actually sound. v6 halved it (23.0 deg, target <10 deg) - progress visible on a scale where on-ground numbers showed nothing.
Change
Suspended-tilt added as a standing acceptance row; run under the honest contact parameters battery (accept_v2 with measured condim 4 / torsional friction 0.035), under which v5 correctly FAILS in agreement with the real robot.
Outcome
The sim battery's verdict on v5 flipped from pass to fail, matching hardware; suspended tilt became the discriminating metric between v5 and v6 (45.9 vs 23.0 deg) when ground metrics differed little.
Mechanism
Contact with the ground closes a stabilizing feedback loop the policy gets for free; removing it exposes the policy's own attitude control authority. A metric measured only in the assisted condition cannot rank policies by the unassisted quantity that hardware will actually demand during perturbations and flight phases.
Applies when
- sim acceptance passes but hardware diverges
- designing an acceptance battery for a legged robot
- two candidates tie on ground metrics
“单点吊那条是最灵敏的失稳探针(拿掉地面这个"免费稳定器"), v5 在地上一切正常却在真机发散, 就是因为验收从没测过这个工况。”
train/WALK_V6_MINIMAL.md § 5. 验收 Teleop fed the sidewalk axis a command beyond its training band - feet clipped; give each axis its own speed setting
teleop-command-band-per-axisGive every command axis its own teleop scale, clamped to that axis's training band, and reproduce any hardware incident in sim with the exact deployed command values before touching training.
Symptom
Robot stepped on its own foot when sidewalking left under teleop - and only when going left.
Context
The teleop tool used one speed setting for all axes: --teleop-speed 0.20 applied to A/D sent cmd_vy = 0.20, above the training band's top (0.08-0.18) where foot-spacing margin is thinnest. Sim reproduction of the incident (product policy, pw0.8, 5 seeds x 20 s, true collision threshold = single foot width 104 mm): at vy 0.20 the minimum foot distance was 111-115 mm - 7-11 mm from self-collision - vs 147 mm at vy 0.10. Left was 4x more dangerous than right (25% vs 6% of time inside the 160 mm soft wall at vy 0.10), matching the left-only symptom; the margin did not degrade over time (pressing more just lengthened exposure).
Change
deploy_policy gained --teleop-side (default 0.10), separating the lateral speed from the forward speed so each axis's teleop command sits inside its own trained band.
Outcome
Command now inside the band with 43 mm margin at default; the incident became a quantified, reproduced, closed account rather than a mystery.
Mechanism
The policy's competence envelope is the training command distribution per axis; teleop mappings that share one scalar across axes silently command out-of-band inputs on the weakest axis. Asymmetric risk (left vs right) came from the policy's own chirality bias, so a symmetric command produced an asymmetric hazard.
Applies when
- wiring a joystick/teleop layer over a learned policy
- a hardware incident occurs on one command direction only
- training bands differ across command axes
“A/D 一直与 W/S 共用速度档，所以按 A 下发的是 vy = 0.20 —— 既超训练带（0.08~0.18）上沿 … 0.20（遥控实际值）| 111~115 mm | 7~11 mm … 且左比右危险 4 倍 … 处置：deploy_policy 新增 --teleop-side（默认 0.10），侧移与前进档分开。”
train/C_LADDER_RUN.md § 3p. 一 向左走踩到自己 → --teleop-speed 0.20 同时喂给了 vy Foot dragging is an attractor, not a low amplitude - and joint damping is the mode switch, adjustable at deploy time
swing-bistability-damping-switchWhen a quality metric is bimodal, stop treating it as an amplitude to be trained up: map the modes against initial conditions and plant parameters, find the parameter that switches basins, apply it first as a deployment lever, and only then bake it into the training distribution (as a plant-family shift, never as an execution-mapping change).
Symptom
s2e_pd-1400's swing height "median 12.1 mm" hid a perfect bimodal distribution: 20 seeds split into a drag mode (2.6-4.9 mm) and a step mode (19.3-24.0 mm) with NOT ONE seed in between - the median sat in the empty gap, and "swing debt -11 mm" really meant "50% probability of falling into the drag attractor".
Context
Two designed experiments closed the mechanism. Test A (nominal plant, 40 seeds): step 42% / drag 58% / middle 0 - at nominal gains, initial conditions alone pick the mode, both modes 100% survivable. Test B (fixed init, kp x kd grid): kd is the mode SWITCH - at kd 1.3 all surviving cells step (13-22 mm), at kd 0.7 nearly all drag (2.7-4.3), only at kd 1.0 does init get a vote; kp >= 1.2 is dangerous (5/6 falls). Global verification at kd x1.3 (20-seed, delay 2): survival 20/20 at ZERO cost, step share 42 -> 80%, swing median 12.1 -> 18.4 mm, slip record low 334, thicker tilt margin - costs: vx 85 -> 78%, saturation +5 pp. A Pareto sweep then priced the knob: step share 42/72/75/88/82/90 across kd 1.00-1.30 with a linear vx tax of -2.3 pp per 0.1 kd - the basin gain is fully collected at kd 1.20 ("1.30 是 over-damping 纯多付税"). Mechanism: low damping leaves a landing micro-oscillation / ground-slide channel the policy can exploit to drag; damping plugs the channel.
Change
Deployment lever adopted: kd-scale 1.20 (conservative 1.15) as the legitimate successor to the power-0.8 crutch ("前者削幅度保稳,后者堵 拖地通道换步态,且不牺牲存活"); training-side prescription: move the DR band to nominal-1.2 x (0.9,1.1) = [1.08,1.32], deleting the [0.7,1.0) drag-teaching zone - a contract-level change requiring digest re-baselining, gated on measuring the real robot's actual kd dispersion first.
Outcome
The kd surgery rung (s2e_kd) delivered basin 8 -> 11/20, slip 405 -> 331, vx 81 -> 85% with no out-of-band fragility (below-band check 20/20) - "拐杖烧进分布的正确姿势", explicitly contrasted with the failed s1g amplitude version: this one changes the plant family the policy has seen, that one changed the execution mapping the policy would have to relearn.
Mechanism
The gait's swing behavior is a bistable dynamical system whose basin boundaries are set by plant parameters; a policy trained across a kd band that includes the drag basin has learned to inhabit it. Shifting the deployed (and then trained) damping moves the system into the step basin without touching the policy - a plant-side fix for what looked like a training deficiency.
Applies when
- a gait quality metric splits into distinct modes across seeds
- deciding between more training and a gain/damping change
- converting a deployment crutch into a training-distribution change
“20-seed 里拖地模式 2.6~4.9mm 与迈步模式 19.3~24.0mm 各半,中间一个不落 … kd 是模式开关——kd1.3 下 6/6 存活格全迈步 … kd0.7 下几乎全拖地 … swing 债的解(至少大半)在部署端阻尼档,不在训练端 … 机理:低阻尼下落脚微振荡/贴地滑给了策略顺势拖行的通道,加阻尼堵之。”
train/README.md § swing 双稳态定性 + kd 部署杠杆 (2026-08-07, 用户设计 Test A/B) Low-friction robustness traced to kd DR bandwidth, not friction training - by digging resolved params across 8 lineages, 3840 cells
kd-bandwidth-mu-law-attributionAttribute capability differences by tabulating every lineage's resolved training params and eliminating zero-variance and non-aligned columns first; never let an eval-side override knob serve as the explanation axis, and never write a mechanism into a law before it survives a targeted test.
Symptom
Lineages differed wildly in low-ground-friction survival, and the intuitive explanation - "some trained ground friction, some didn't" - was about to steer the ladder toward a ground-mu training rung.
Context
The attribution ran as a full parameter-vs-result cross: 8 lineages x 4 eval kd levels x 6 mu levels x 20 seeds = 3840 cells, with each lineage's RESOLVED training params dug out and compared item by item. First kill: all 8 lineages had ground mu pinned at (1.0,1.0) - zero variance - so low-mu differences cannot come from friction training at all. The only training parameter aligned with the mu score was kd DR bandwidth: narrow (<=0.24) lineages scored 19.9/19.5/19.5, wide (>=0.40) scored 17.1/15.2/14.6/14.2/12.8 - the two groups completely non-overlapping. Every rival was excluded item by item (kd center no; kp band no; COM small-beneficial non-driving; friction rung a clean double null 19.5->19.5 and 15.2->14.6; iteration count non-monotonic), and the one clean single-variable causal link confirmed it: the s2e-3 kd surgery (0.7,1.3)->(1.08,1.32) moved the score 17.1->19.5. Counter-proof against "each best at its own operating point": the narrow-band lineage evaluated OUT of band (18.2) still beat the wide-band lineage at its own band center (9.2). Two axes were ordered never to be conflated (the first attribution's own error): training kd bandwidth is a parameter axis / lineage property; the eval-side --kd-scale knob is a plant axis (more damping physically helps on slippery floors for ALL policies) - "plant 轴只能当部署缓解,不能当 归因". A tempting mechanism story ("drag vs step attractor") was tested and falsified, and explicitly kept OUT of the law: "机制未定, 不入定律".
Change
The planned ground-mu training rung was recommended closed ("建议 不开") in favor of a kd band-narrowing rung (0.8,1.2)->(0.9,1.1) centered on the deployed value - with a pre-registered risk that the law demands "bandwidth = measured dispersion" and the real robot's kd dispersion was not yet measured; if it exceeds +/-10%, narrowing sacrifices real coverage and the rung must yield.
Outcome
A whole training rung was deleted from the ladder by attribution alone (the second S2 pass dropped mu and push, 5 rungs -> 3); floor material became a deployment-selection input (mu <~0.6 -> deploy the kd1.2 gain profile) rather than a training target.
Mechanism
Cross-lineage performance differences must be attributed over the actual training-parameter table, not over eval knobs or plausible stories: eval knobs act on the plant for every policy (a physical effect), while lineage properties come only from training-time parameters. Zero-variance columns are free eliminations, and one clean single-variable rung is worth more than any correlation.
Applies when
- explaining why lineages differ on a robustness axis
- an eval-side knob (gain scale, power) changes results and invites misattribution
- deciding whether to open a DR rung for an axis never actually varied in training
“8 血统地面 μ 训练带全部钉 (1.0,1.0) 零方差,低 μ 差异与「训没训地面摩擦」无关,是 kd DR 带宽的副产物 … 宽 ≤0.24 → 19.9/19.5/19.5;宽 ≥0.40 → 17.1/15.2/14.6/14.2/12.8, 两组完全不重叠。… 训练 kd 带宽 = 参数轴/血统属性;评测部署 --kd-scale = plant 轴 … plant 轴只能当部署缓解, 不能当归因。… 机制未定, 不入定律。”
train/OMNI_V0_SPEC.md § 4. 地面 μ 鲁棒性 = kd DR 带宽的副产物 (2026-08-08) An auto-curriculum ratchet capped out at iter 248 and never engaged - stage difficulty manually or verify engagement
auto-curriculum-engagement-checkPrefer manually staged difficulty with gated transitions; if you use an automatic curriculum, instrument its internal state and alarm when it stops engaging - a saturated curriculum is constant DR wearing a curriculum's name.
Symptom
A curriculum mechanism intended to grow difficulty adaptively (s1f's ratchet) hit its cap at iteration 248 and never bit again - for 96% of the run its effect was equivalent to constant DR, i.e. the curriculum existed in name only.
Context
When external advice suggested graded wz bands (start ±0.15, then ±0.30), the team agreed with grading but explicitly rejected automatic curriculum, citing the s1f episode. The same logic had already been paid for with push grading: ±0.6 failed twice, ±0.3 was feasible - grading matters, but the grade transitions were made by hand at verified checkpoints.
Change
Ladder policy: difficulty staged manually, one band per rung, each transition gated by the acceptance battery; automatic ratchets not used unless their engagement is monitored and demonstrated.
Outcome
Every C-ladder band change (wz ±0.12-0.25 first, wider later) was an explicit, attributable rung; no silent constant-DR-in-disguise runs recurred.
Mechanism
Adaptive curricula couple their own state machine to noisy training metrics; a ratchet that saturates early stops adapting but keeps its name, so the operator believes difficulty is progressing when it is frozen. Manual staging costs more decisions but each decision is observable and reversible.
Applies when
- choosing between auto-curriculum and staged bands for a new skill
- a curriculum's difficulty parameter plateaus early in training
- post-hoc attribution of what difficulty a lineage actually saw
“C2 的 wz 分级（±0.15 → ±0.30，不要一上来 ±0.6）—— 与我们付过学费的 push 分级同形（±0.6 两轮 FAIL，±0.3 才可行）。但必须手动分级，不做自动课程 —— s1f 的课程化棘轮 iter 248 封顶未咬合，96% 时长等价常量 DR”
train/C_LADDER_RUN.md § 1. 采纳 3 条 (C2 的 wz 分级) Fall recovery was defined as the whole chain - any fallen pose, a stable stand, a clean hand-back to walking - and built as a second policy behind a deploy-side switch, not folded into the walking PPO
recovery-two-policies-and-a-state-machineDefine a recovery skill by the whole chain it must complete, including the hand-back to the next controller; if it is built as a separate policy, make the switching logic and its handoff contract a deliverable of their own, and keep the recovery observation contract a subset of the locomotion one so a unified policy stays possible later.
Symptom
A walking robot that falls needs a human to stand it back up. The design question on 2026-08-09 was whether to teach getting up inside the existing omni walking policy or beside it.
Context
The user set the goal as "any fallen pose -> stand up alone -> stand stably", and the spec named the real difficulty as the full chain fall -> recovery -> stable stand -> correctly initialised walking history and clock -> walking, making the deploy state machine a first-class deliverable. A unified single policy had a real-robot precedent (arXiv:2605.18611, a state-dependent gate near 37 deg tilt) but was deferred until a recovery policy and an omni policy were each reliable. The line ran on its own branch and worktree with every walk/stand/omni/run config path untouched. The development path copied the G1 learned get-up logic (arXiv:2502.12152): first find any feasible get-up (ugly accepted), then add smoothing, torque and real-robot constraints. The recovery contract kept the base 45-dim observation (command slice held at 0, no gait phase, no frame history - their reasons do not apply to a skill without a clock or a velocity task), so it stays a prefix of the 215-dim omni contract and a later merge is not foreclosed.
Change
Two policies and a deploy-side switch instead of one retrained walking policy; recovery got its own minimal contract (45 dims, full-range action, later the beta-anchored profile) and its own acceptance battery.
Outcome
The split held for the whole line: on 08-14 deploy_policy gained a second (PolicyIO, ONNX) pair behind --recovery-policy, each loaded under its own manifest contract, and the runbook runs stand_v1b or omni_c4_ff800 as the locomotion side with recovery_v3_1p1c. The literature scan of 08-10 found that every verified get-up implementation deploys one end-to-end policy (or softly gated experts) and stages only on the training side - so the runtime state machine here is the walk/recovery switch, not a staged get-up.
Mechanism
A separate policy keeps each reward table single-purpose and lets a proven walking lineage stay byte-frozen; the cost moves to the handoff, where every piece of state one policy leaves behind (history, clock, last action, command) must be reset for the other.
Applies when
- adding fall recovery or get-up to a robot that already walks
- choosing between one unified policy and a switched pair of policies
- designing the observation/action contract of a secondary skill
“先做 recovery policy + omni policy 两个策略,部署侧状态机切换;不把 recovery 硬塞进现有 omni PPO。 … 任务定义:**任意跌倒姿态 → 自己站起来 → 稳定站立**。真正的难点不只是"起身", … omni walk**(§6 部署状态机是本 spec 的一等公民,不是附录)”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §0 目标口径与架构判决(用户 2026-08-09 定) When the training reset distribution changes, freeze the acceptance distribution separately and pin its seed - the same checkpoint measured twice differed by 3.8 and 10.2 points
frozen-acceptance-distribution-and-pinned-seedAcceptance distributions are frozen artifacts, decoupled from whatever the training distribution becomes and evaluated with a pinned seed, and every gate row carries its sample size so that a small-n row is never read as a regression.
Symptom
R0.3 added easier roll-arc start states to the training resets, and the acceptance script shared the training category table; separately, the same checkpoint scored side 3.8 points and mid 10.2 points differently on two runs of the same script.
Context
Acceptance ran 512 parallel envs drawn from the fall categories. Had it kept following the training table, 15% of acceptance samples would have landed on states easier than prone - inflated scores and generations that could not be compared. The two same-checkpoint runs had identical per-item height medians, so the ruler had not changed; the spread was reset resampling noise (side n ~ 169, sigma 2.3%; mid n ~ 35, sigma 8.3%). The spec's "20 seeds" had always meant controlled seeds.
Change
accept_recovery.ACCEPT_CATEGORIES pinned to the four R0-R0.2 categories and decoupled from the training FALL_CATEGORIES; --seed 20260809 pinned, after which two consecutive runs were bit-identical. The mid row (n ~ 35) was labelled the bluntest gate.
Outcome
Every later generation (R0.3 through V3.1) was scored on the frozen distribution and seed, which is what let R0.3's intermediate state be read as "no measurable gain" (62.7 -> 62.1%) and R0.2's mid drop be booked as noise rather than a regression.
Mechanism
An acceptance set that follows the training distribution measures a moving target, and an unpinned reset draw adds sampling noise that small-n rows cannot absorb.
Applies when
- the training reset or command distribution changes between generations
- repeated evaluations of one checkpoint disagree
- a small category drives a pass/fail decision
“**分布冻结**:`accept_recovery.ACCEPT_CATEGORIES` 钉死 §5 四类 … 与训练侧 `FALL_CATEGORIES` **解耦**。 … **side 差 3.8 点、mid 差 10.2 点**(h 中位逐项一致,证明不是尺子变了)—— 纯 reset 重采样噪声 … 钉死后两次连跑逐位相同。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §14 验收尺子的两处加固(R0.3 起生效,向后兼容) A reward on a quantity the actor cannot observe teaches "produce less of it", never "correct it" - closed-loop correction needs an outer loop
reward-observability-limitBefore adding a reward, check the actor can observe (or infer) the quantity: unobservable-error rewards buy only average suppression - route correction tasks to an outer loop whose commands stay in distribution, and do not break a frozen contract to add an observation a deploy-side loop can supply.
Symptom
Heading kept drifting despite world-frame yaw rewards, and a reviewer proposed heading-error rewards - raising the question of what yaw shaping can even teach this actor.
Context
The adopted architectural verdict: the actor's 45-dim base observation cannot see accumulated heading at all - projected_gravity is invariant to rotation about the gravity axis, and omega_z is a rate, not an angle. World-frame yaw-rate rewards are therefore privileged shaping that can only teach "少产生旋转" (generate less rotation), never "偏了以后拉回原线" (pull back to the line after drifting) - the policy cannot represent the error it would need to correct. The S1 gate (<=5 deg / 10 s) demands exactly the former, so the stack is right for its gate; active heading correction is assigned to the deployment outer loop (--heading P-loop converting heading error into in-distribution wz commands) plus small-wz training - and the 215-dim contract is explicitly NOT extended with a heading observation ("契约不加 heading 观测,冻结不动"). The reviewer's companion bias hypothesis was adjudicated with data: drift is bimodal - a basin mechanism decides whether you leave (seeds vary +/-16-46 deg vs -385 to -391 deg), and once out, rotation direction is constant (weight chirality; candidate root: the phase clock always swings left first).
Change
Yaw shaping kept as rate-tracking (three-layer stack); heading correction owned by the deploy outer loop; contract frozen; the "which behaviors need an outer loop" question settled by observability analysis rather than reward tuning.
Outcome
Stopped a contract change and a futile reward direction; drift work split correctly into rate-suppression (trainable) and error correction (outer loop), consistent with the earlier measured 10x drift reduction from the deploy-side loop.
Mechanism
A policy can only condition on its observation sigma-algebra; rewards on functions outside it shift the marginal action distribution (open-loop average effects) but cannot create feedback on the unobserved variable. Whether to add an observation, an outer loop, or accept average-shaping is decided by the task's gate: suppression gates need shaping, correction gates need the variable in some loop's view.
Applies when
- adding rewards on accumulated/世界-frame quantities (heading, position)
- deciding between a new observation, an outer loop, and shaping
- a drift symptom persists across reward-weight changes
“actor 的 45 维基座观测不到累计航向(projected_gravity 对绕重力轴旋转不变,ωz 是速率不是角度)——世界系 yaw 奖励是特权塑形,只能教「少产生旋转」,不能教「偏了以后拉回原线」。… 主动纠偏闭环 = S3 把小 wz 进分布 + deploy --heading 外环 … 215 契约不加 heading 观测,冻结不动。”
train/OMNI_V0_SPEC.md § 3. 评审④判决(2026-08-06,S1.3 开训前) Symmetrizing the config made the gait MORE asymmetric - the asymmetry lived in the policy weights
asymmetry-in-weights-not-configLocalize a persistent asymmetry by intervening at the config layer first: if the symptom survives (or worsens), it is in the weights - fix it with symmetry-constrained training, not with trims or offsets.
Symptom
walk_v1 on hardware: straight-line command curved 149 deg in 15 s (9.9 deg/s) with 3.06 m lateral runout; turn gain +31% one way vs +129% the other (75% difference); knee asymmetry 4.4 deg in sim, 9.6 deg on the robot.
Context
The obvious suspect was the asymmetric default pose in the config. The decisive test: symmetrize standing_pose and run the SAME policy in sim - the asymmetry got LARGER (hip_pitch 6.8 -> 9.3 deg). Root cause therefore not in the config but baked into the policy weights: PPO without a symmetry constraint commonly converges one-sided, because splitting the work 50/50 and loading one side yield the same return, and the gradient falls randomly into one of the equivalent optima.
Change
Fix redirected from config trimming to retraining with mirror data augmentation (walk_v2 spec) - a weights-level fix for a weights-level disease.
Outcome
With augmentation (and the symmetric-default precondition), stand_v1 reached 0.0 deg asymmetry on all six joint pairs (from 4.4-7.7 deg), height fluctuation 7 mm -> 1 mm, mean |action| down 33%.
Mechanism
Reward-equivalent solution families (who carries the load) leave the symmetric solution unpreferred; SGD picks an arbitrary member and entrenches it. Config changes move the coordinate frame around the entrenched asymmetric function - they cannot move the function. The counterfactual test (change config, watch symptom) localizes the layer the disease lives in.
Applies when
- a robot veers or loads one side despite a symmetric-looking config
- deciding between config trims and retraining for an asymmetry
- mirrored-turn gains differ by tens of percent
“根因不在配置里:把 standing_pose 对称化后在 sim 里跑同一策略,不对称反而变大(hip_pitch 6.8°→9.3°)—— 说明不对称烙在策略权重里。这是无对称约束的 PPO 的常见收敛结果(左右各担一半与一边多担的回报相同,梯度会随机落进其中一个)。”
train/RETRAIN_v2.md § 1. 为什么是对称增强(证据) Two machines, one configuration - every gain, offset and torque limit changes only in robot.yaml, whoever edits pushes at once, both checkouts show the same commit before the robot moves, and a pulled policy file is size-checked and synced before power-off
two-machine-config-disciplineTreat the robot's configuration as a versioned artifact with one source of truth, push every change immediately, verify identical commits on every machine before a hardware session, keep hardware limits in the repo and the firmware in sync in both directions, and verify transferred model files (size, digest) before running them.
Symptom
The robot's onboard computer runs the bridge and deploy scripts from its own checkout while training and analysis happen on other machines; a hot fix left on one side, or a half-written file, silently makes the robot run something other than what was evaluated.
Context
The operator runbook's "wall version" of the two-machine discipline: configuration changes only in robot.yaml (calibration offset/sign, gains, torque limits), committed and pushed from the Mac, pulled on the robot, bridge restarted; code is not edited on the robot, and if it is, it is committed and pushed on the spot - nothing unpushed overnight; 30 seconds before every real-robot session both checkouts must show a clean status and the same last commit hash; changing tau_max requires writing the motor's limit_torque too (and the reverse); re-zeroed motors require re-measuring offsets. The recovery line added: after pulling on the robot, check the ONNX is not zero bytes (a lesson from a corruption incident on 08-12) and sync before cutting power; the recovery and main lines are separate worktrees, each pulled with --ff-only.
Change
Operating rules, pinned on the wall and repeated in the hanging checklists ("git pull, both machines on the same commit").
Outcome
The sources record the rules and the incident that produced the size check; they do not record a count of sessions the rules caught.
Mechanism
A policy is evaluated against one configuration; any divergence between the machines, or a truncated file, turns a hardware result into a result about an unknown configuration.
Applies when
- a robot's onboard computer and a workstation both hold the configuration
- someone hot-fixes code or gains on the robot
- model files are copied or pulled to the robot before a session
“改配置只改 robot.yaml(标定 offset/sign、增益、限扭全在里面)→ Mac git commit + push → NX git pull → 重启桥。 … 谁改完谁立刻推,永远不留未推送的改动过夜。 … 每次上真机前 30 秒检查:两边 git status 干净、git log -1 哈希一致。 … 铁律不变:改 tau_max 必须同步写电机 limit_torque(反之亦然);电机重新标零后 offset 必须重测回填 yaml。”
RL系统/FOLLOW THIS copy 2.md § ② 双机维护纪律(贴墙版) Verify changes in the run's resolved config (and checkpoint md5), never in the source you edited
resolved-config-is-source-of-truthAttribution and single-variable claims must be made on the resolved per-run config (and checkpoint hashes), not on source diffs; verify every intended variable landed before burning compute, and verify every rollback byte-level against the historical resolved config.
Symptom
An intended arm-B config change never reached the training run - the run was grid-identical (117/117 cells) to its C2 predecessor - and the burn was only understood afterwards.
Context
The repo's discipline hardened around the logged resolved config (logs/<run>/params/env.yaml) as the only source of truth: (1) the C2 root-cause analysis was performed against the checkpoint's logged env.yaml, not the code ("以真相源 23-19-25/params/env.yaml 核实"); (2) C4 added a pre-flight: grep the landed env.yaml for the new keys, and compare the first checkpoints of the two arms - identical md5 means the variable did not land, stop immediately; (3) the C4 full rollback was accepted only after starting a 1-iter run and byte-comparing its resolved env.yaml against the historical 700-era file (identical except 4 dormant schema fields, each verified to be at its no-op default).
Change
Standing pre-flight and post-change verification: dump/diff the resolved config that the run actually consumed; use checkpoint hash equality as a cheap "variable landed" detector between arms.
Outcome
Caught the not-landed variable class of failure; made the rollback provably equivalent to the historical training state rather than believed-equivalent.
Mechanism
Between edited source and the running experiment sit layered overrides, env-var switches, and registration logic; only the resolved, serialized config reflects their composition. Diffing at that level tests the actual experiment; diffing source tests intent.
Applies when
- launching an A/B pair or any single-variable rung
- rolling back to a historical training state
- a run behaves as if a change was never applied
“开训前先验落盘 cfg（上一轮臂B 的改动没进 run，与 C2 逐格 117/117 相同）：grep -E "base_com|joint_friction|push_robot|track_lin_vel_y_exp" logs/<run>/params/env.yaml 另：两臂第一个 checkpoint 的 md5 若相同 = 变量没进去，立刻停。”
train/C_LADDER_RUN.md § 3d. ⚠️ 开训前先验落盘 cfg / 3l. 回退清单（验证） A walking policy's tilt cutoff is a legal state for a recovery policy - the default 45 deg fall guard had to be raised for recovery tests and is disabled once the switch owns falls, so the abort chain becomes the recovery timeout, the operator's cut, and the firmware torque limits
fall-guard-becomes-a-stateWhen a new skill makes a safety cutoff's trigger a legal state, replace the cutoff with a bound of the skill's own (a timeout ending in a safe stop) instead of just switching it off; keep the operator's cut and the firmware limits as independent layers, and write every flag change into the run sheet.
Symptom
deploy_policy's default protection stops the robot beyond 45 deg of tilt. A recovery policy starts lying at roughly 90-97 deg, so under the default it is refused on the spot - a flag the first hanging checklist forgot.
Context
The layers in the sources: deploy_policy's tilt cutoff (default 45 deg, a line in the safety chain); for standalone recovery tests the cutoff was raised (110 deg in the spec's A/B sheet; 181 deg, effectively off, in some runbook commands); with --recovery-policy the cutoff is disabled because a fall is now a state, not an exception, and RECOVERY lasting over 15 s ends in a safe stop (the runbook calls it the line where the spotter steps in). Independent of the policy: firmware torque limits checked at start (12/17/11 N*m, set_torque --check), the operator cutting enable at any kicking or oscillation, and in the one-leg teleop a space-bar stop that puts the foot down.
Change
The flag was added to the run sheets, and the FSM replaced the removed cutoff with its own bound (the timeout).
Outcome
The spec records the flag omission and its fix; it does not record the FSM's timeout being exercised on hardware.
Mechanism
A safety cutoff encodes one policy's notion of "abnormal"; a new skill whose normal operation lies beyond it either cannot run or runs with the cutoff off, and only a replacement bound keeps the chain closed.
Applies when
- deploying recovery, fall-damage or acrobatic skills behind existing safety checks
- a run sheet disables a protection flag
- listing the abort chain for a hardware session
“--max-tilt-deg(默认 45°,安全链第 13 行写的那个)。recovery 的合法状态覆盖整个倾角域,把它抬到 181 = 实效关闭 … RECOVERY 超时 15s 会自动安全停(看护介入线)”
RL系统/FOLLOW THIS copy 2.md § FSM 吊挂首测 ② 落地测 / #### Recovery Policy (operator runbook, undated) Every rung gets written stop criteria - hit any one, stop; tightened when priors say results should come fast
preregistered-stop-criteria-per-rungFreeze per-rung stop criteria (old-skill floors, new-skill deadline, oscillation signature, known pathology signatures) before training, stop on first hit - and shorten the deadline in proportion to how fast your mechanism says results should appear.
Symptom
Continued training past the point of degradation had previously destroyed capabilities (C1 trained away the root's backward skill); without hard stop rules, sunk-cost reasoning keeps runs alive too long.
Context
Standard rung protocol - eval_c_matrix every 100 iters at 5 seeds with a fixed criteria list, frozen before training ("开训前写死，事后不许改"): (1) stand or vx+0.15 at <=3/5 for two consecutive points; (2) vx+0.15 tracking <70% for two consecutive points (vs recorded root baseline); (3) oscillation signature - survival repeatedly crossing zero between adjacent checkpoints, stop on first occurrence, do not wait for confirmation; (4) new-skill-no-progress deadline (e.g. wz tracking <30% or still same-signed after iter 400); (5) a known lineage pathology signature (s1c arm-B: scatter -> half-recover -> collapse).
Change
Stop budgets scale with prior knowledge: when C4-redo4's feed-forward was already proven open-loop, the no-progress deadline was tightened from +500 to +100 iters ("前馈是开环就给 71~96% 的 … 若 +100 还没有 … 不值得再烧").
Outcome
Multiple rungs were stopped exactly on criterion (C4-redo criterion 4 at iter 1200; redo2/redo3 at absolute 900), converting each into a clean hypothesis test instead of a drifting run; no rung burned its full cap on a dead hypothesis.
Mechanism
Degradation of retained skills is a lineage-level injury that later training often cannot undo, so detection latency is capital loss; and a stop rule written before the run cannot be bent by hope. Tightening deadlines when priors predict immediate results converts "no result yet" into evidence against the mechanism.
Applies when
- starting any resumed/curriculum training rung
- deciding whether to keep training a run that shows early regression
- a mechanism-backed change should produce results immediately
“每 100 iter eval_c_matrix.py --seeds 5，命中任一即停 … 存活在相邻 checkpoint 间反复跨越 0（出现即停）… 收紧版：绝对 iter 900（= +200）时 vy±0.10 跟踪仍 <30% 即停”
train/C_LADDER_RUN.md § 3h. 预注册停梯判据（开训前写死，事后不许改） The recovery line's real-robot verdicts live in three places that disagree - the spec's "fairly stable, first real stand-up" for v2_6, an undated runbook note that only v3_1p1c works, and a first run whose details were never recorded
write-hardware-verdicts-backA hardware verdict is a dated entry in the authoritative ledger - policy file and stamp, gain profile, floor, battery, tries, log file, what was seen - written back the same day; a note in a command file is a pointer, not a verdict, and a newer verdict that contradicts an older one must say so.
Symptom
Asked "which recovery policy works on the real robot", the sources give different answers, and none of them carries the conditions of the test.
Context
08-09: the first real run was stopped as violent and dangerous; which ONNX, which gain profile and whether a log existed were marked "to be recorded" and never were. 08-11: v2_6 was "fairly stable", the line's first real get-up, with splits after standing; v2_5b's result and both CSVs were "to be reported". 08-14: v3_1p1c was stamped and pushed, with "the real first test still needs the user present"; the spec records no hardware result for it. The operator runbook (undated) puts above the v2_6 and v2_5b floor commands the note that none of the recovery policies below work, only recovery_v3_1p1c - a verdict never written back into the spec, with no date, floor, battery, number of tries or log attached.
Change
None recorded in the sources; this card records the gap.
Outcome
The line's authoritative record ends with v3_1p1c as the product awaiting its first real test, while the operator's note implies it is the only one that works and that v2_6 (recorded as a success) does not.
Mechanism
Verdicts given at the robot travel by word of mouth and command-file comments; without a record carrying the conditions, a later reader cannot tell a changed verdict from a changed floor, battery or stack.
Conflicts
§43 (2026-08-11) records v2_6 as the first successful real get-up ("fairly stable"); the undated runbook says every recovery policy except recovery_v3_1p1c does not work; §49 (2026-08-14) says v3_1p1c's first real test was still pending. The runbook's claim has no date and was never written back to the spec, so it cannot be ordered against §43.
Applies when
- choosing which policy to deploy from an operator's notes
- a hardware session ends without a written result
- two documents disagree about what worked on the robot
“下面的recovery都不行 只有recovery_v3_1p1c.onnx”
RL系统/FOLLOW THIS copy 2.md § #### Recovery Policy (operator runbook, undated) Training-log reward values and fixed-command eval values live on different distributions - comparing them once claimed a 44% improvement that was really 6-10%
same-distribution-reward-comparisonQuote reward-term values only with their distribution attached (command range, DR on/off, environment), and compare across runs only when those match; re-measure in a common environment before claiming any improvement percentage.
Symptom
A v6-era analysis concluded slip had dropped 44% by comparing the training log's Episode_Reward against values calibrated in a fixed-command play environment; a same-condition re-measurement showed the true improvement was 6-10%.
Context
The training log's reward is an expectation over the training command distribution (vx 0.15-0.5, yaw +/-0.6, with pushes and domain randomization); play-environment calibrations are taken at a single fixed command with DR off. Subtracting one from the other compares apples to oranges - the warning was written into the v7 pre-flight: "奖励数值只能在同一指令分布下比较 … 拿它和固定 cmd 的 play 环境标定值相减会得出错误结论(v6 那轮已经栽过一次)".
Change
Rule adopted: any before/after reward-term comparison must hold the command distribution, DR state, and evaluation environment fixed; training-log values compare only against training-log values of runs with identical command/DR configs.
Outcome
The phantom 44% improvement was retracted; later term-level accounting (e.g. the C4 ignore-floor work) consistently specified its distribution before quoting numbers.
Mechanism
A reward term's expectation depends on the visited-state distribution as much as on the policy; changing the command distribution or DR moves every term's baseline. Cross-distribution differences therefore measure the distributions, not the policy change.
Applies when
- comparing reward telemetry across training runs or vs play evals
- claiming improvement percentages from training logs
- term-level reward accounting for diagnosis
“奖励数值只能在同一指令分布下比较。训练日志的 Episode_Reward 是在训练指令分布上算的(vx 0.15~0.5 / 偏航 ±0.6 / 带推力与域随机化), 拿它和固定 cmd 的 play 环境标定值相减会得出错误结论(v6 那轮已经栽过一次: 据此以为滑移降了 44%, 同条件对拍只有 6~10%)。”
train/WALK_V7_SPEC.md § 3. 开训自查 ⚠️ The fallen-state reset was designed, not sampled from SO(3) - fixed category shares with jitter, a low drop that settles physically, equal left/right shares for mirror augmentation, and a numeric check before training
fallen-pose-reset-distributionBuild a fallen-start distribution from named, physically plausible categories with jitter and a settle phase, keep mirrored categories at equal probability, and check the realized shares and penetration numerically before spending a training run on it.
Symptom
A get-up policy can only learn from the fallen states its resets produce; uniformly random orientations produce ground-penetrating and limit-jammed states the robot can never be in.
Context
R0 reset_root_fallen: supine 30%, prone 30%, side_l 15%, side_r 15%, mid (random axis 50-125 deg) 10%, +/-15 deg jitter, full yaw, dropped from 0.28-0.40 m and left to settle under physics, joints uniform inside the soft limits with a 5% margin plus small random velocities. Random SO(3) was rejected (the advisor agreed). side_l and side_r must have equal probability because mirror augmentation turns a left fall into a right fall. The advisor had proposed supine and prone only for R0; the spec included side and mid because the feasibility accounts showed physical solutions for all of them, and wrote "narrow back to supine+prone" down as the first fallback. With no display on the training box the reset was checked numerically instead of by eye.
Change
Category mix as above; realized shares, settle height and penetration measured over 512 envs before the first run. A fallen-state bank (real falls, settled and stored) was pre-registered for R2.
Outcome
Realized shares 29.3/31.6/16.4/17.8% against the config, settle +0.262 m, final penetration 0/512 (a 0.10 m peak at the write instant, ankle links only, pushed out within 80 ms because the 0.28 m drop floor is shorter than a fully extended leg). R0's failure was a reward basin, not a reset artifact. The fallen-state bank stayed unbuilt through V3.1 (checklist item open); R0.3 later re-sliced the prone share into roll_l/roll_r bands, which is what forced the acceptance distribution to be frozen separately.
Mechanism
A category-structured, physically settled start distribution keeps training on states the robot can actually occupy, and equal mirrored shares keep mirror augmentation a pure doubling of data rather than a bias.
Applies when
- designing reset distributions for get-up, recovery or multi-contact skills
- mirror/symmetry augmentation is on and the task has chiral start states
- no viewport is available to inspect resets on the training machine
“角度 jitter ±15°、yaw 全域、0.28~0.40 m 低空放下由物理沉降,关节软限位内 均匀(留 5% 余量)+ 小随机速度。**不用 random SO(3)**(会采出穿地/极限卡死 等现实不可能状态,顾问同判) … side_l/side_r **概率必须相等**(镜像增强的样本同分布前提)”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §3 R0 任务定义 / §9 核查单 Before training a one-leg stand the spec named the cheapest cheats - hopping on the support foot, a raised foot resting unloaded, a leg tripod - and gave each a countermeasure and a gate; one still appeared and was caught by exactly those gates
enumerate-cheapest-cheats-before-trainingBefore training, list the cheapest behaviours that would satisfy each reward term without doing the task, give each a countermeasure in the reward and a gate in acceptance, and prove the intended behaviour is reachable with a probe - then treat any gate the policy games as evidence about the reward, not the gate.
Symptom
The literature's single-leg benchmark reports eight state-of-the-art general policies holding a clean one-leg stand 0 times out of 90 - they survive by sneaking steps and hops - so the task's first adversary was the policy's own cheating.
Context
The spec's shape self-check ("what is the zero-cost option?") listed, for the one-foot bucket: the cheapest cheat, a foot resting on the ground without load, countered by a 5 N contact threshold plus positive swing income; the second cheapest, small hops on the support foot to reset balance, countered by a continuous support-air penalty plus a gate of zero support-foot flight segments. The probe that preceded training had already seen a third: early low-lift postures "survived" by pressing the swing foot at 78-95 N, a leg tripod, removed by folding the shank back. The two-foot bucket was checked too: its zero-cost behaviour is ordinary standing, with no odd base state.
Change
Countermeasures and gates written before training: swing-contact and support-air penalties, gate 2 (zero swing-foot contact frames above 5 N), gate 3 (zero support-foot flight segments).
Outcome
The first run still found the unloaded-foot cheat (a binary reward band gave it no gradient to lift) - and it was caught, by the contact gates and the cross-simulator comparison, not discovered on hardware. The retrained V0 passed all gates 40/40, including zero support-foot flight after the flight detector was corrected.
Mechanism
A policy optimizes the reward, not the intent; the cheapest behaviours that satisfy the reward are predictable from the reward's structure, and a gate written for each before training turns a silent cheat into a failed row.
Applies when
- designing rewards for balance, contact or "hold still" tasks
- benchmark policies are known to cheat the task
- writing acceptance gates for a new skill
“文献里 8 个 SOTA 通用策略在单脚站基准上 0/90 干净保持， 全靠偷步偷跳活命，这是本任务的第一反作弊对象 … 单脚桶下最便宜的作弊是"脚虚放地上不受力"——接触判定 >5 N 力阈（沿用），配 swing_height_band 正收入拉开。 … 第二便宜是"支撑脚小跳重置"——support_air_penalty 连续罚 + 验收门支撑脚腾空段=0 双保险。”
git:Lucen V2@origin/oneleg-line:train/ONELEG_V0_SPEC.md § §0 目标口径 / §5 形状自检(零成本选项是什么) A plateau in a training curve was a population mix, not a half-learned skill - 60% standing at 0.372 m and 40% sitting at 0.19 m - and a category at hard zero stayed at zero through 3,000 more iterations
zero-partial-credit-is-not-an-iteration-problemBefore buying iterations for a plateau, split the metric by category and check whether it is bimodal; a category at hard zero with no partial credit is missing a capability or a reachable state, and more iterations under an unchanged config will only polish the categories that already work.
Symptom
After R0.1 the training-side base_height sat near 0.29 m and the curve was still climbing at the iteration cap, which read as "train it longer".
Context
Candidate A (user decision) was a child-run from R0.1's last checkpoint with zero config change - the logged env.yaml files differ only in log_dir - for 3,000 more iterations (R0.2). The per-category acceptance split was already available: prone had scored 0/156 with no partial credit.
Change
Continue training unchanged, then read the result by category rather than by the pooled curve.
Outcome
supine 91.5 -> 96.4%, side 82.2 -> 87.9%, get-up 0.96 -> 0.84 s, pose error 0.91 -> 0.54 - all improvements to categories that already stood. Prone stayed 0/153; mid 55.9 -> 41.2% was within noise (n=34). Height by category was binary - standing groups 0.372/0.373 m, seated groups 0.187/0.194 m, nothing between - so the pooled 0.29 was 0.61 x 0.372 + 0.39 x 0.19 = 0.30 (measured 0.307): six in ten standing, four in ten sitting. The rendered prone episode was still kneel-sitting at t = 8 s.
Mechanism
The pooled mean of a binary outcome only moves when the mix moves; PPO kept polishing the subpopulation that already succeeded while the failing one produced no advantage signal to follow.
Conflicts
R0.2 recorded the missing capability as "prone lacks rolling over"; R0.3's end-state confusion matrix retracted that - prone had righted its torso in 159/159 episodes and was failing to stand from the W-sit. The lesson that iterations could not fix it holds; the named cause was wrong.
Applies when
- a training curve plateaus while acceptance shows one category at zero
- deciding between "train longer" and "change something"
- pooled training metrics are read as the typical episode
“**分类别 h 中位把"平台 = 人口混合"钉死了**:数值是**二值**的 —— 站立组 0.372/0.373,坐姿组 0.187/0.194,**中间没有过渡态**。 … **这也是本仓此后读该指标的通用告诫:全体混合的期望会把 双峰分布平均成一个不存在的中间值,必须分类别看。** … **结论:A 不能过门,原因确定为 prone 缺"翻身"这一技能,不是迭代不够。**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §13 R0.2(recovery_r0_2,child-run 续训):A 走完了 —— 推不动 prone When a contract default changes, old policies must run under a pinned legacy profile - a silent clock swap is out-of-distribution on hardware
legacy-profile-pinningTreat every trained policy as bound to the contract values of its training era: version the deployment profiles, pin old policies to their era's profile in every command, and never let a changed default silently apply to an old artifact.
Symptom
The walk profile's gait clock moved from 0.40 s to 0.50 s for new training, but versions v5-v9 were all trained at 0.40 s - running them under the updated default would silently feed a 25% slower phase clock to policies that never saw one.
Context
The re-test runbook hard-codes --policy-profile legacy_walk_040 into every command for the old versions, with the warning not to omit the flag: the mismatch is invisible (no error, no crash) but puts the policy out of distribution on hardware, where the same file had already documented that off-clock operation collapses gait quality.
Change
Deployment profiles versioned per training era; historical policies permanently associated with their era's profile; runbooks write the profile flag explicitly rather than relying on defaults.
Outcome
Old policies stayed runnable and comparable after the contract moved on; the silent-mismatch failure mode was closed by convention.
Mechanism
Changing a shared default rebinds every old artifact to a contract it was not trained under; unlike a schema break, a value change produces no error - only degraded, unexplainable behavior. Version-pinned profiles make the binding explicit and permanent.
Applies when
- changing any default in the deployment contract (clock, scales, gains) while old policies remain in use
- writing runbooks that mix policy generations
- a re-tested old policy behaves worse than its era's records
“2026-08-02 起 walk profile 的时钟改为 0.50(WALK_V10_SPEC §3)。v5~v9 全是 0.40 训的,本文件所有命令已改带 --policy-profile legacy_walk_040 ——不要省掉这个 flag,否则是拿慢 25% 的相位时钟静默喂旧策略(分布外,真机危险)。”
train/REAL_SWEEP_V5_V8.md § 1. 预检 ⚠️ 时钟改为 0.50 The walking lines' safety setting, power-scale 0.8, broke the recovery policy's full-range contract - it cut the ends of the joint travel (4/50 could not get up) and left the torque spikes untouched; a kp x 0.9 gain profile inside the trained kp band did the job
power-derating-cuts-full-range-contractA deployment derating knob means something only relative to the action contract: before reusing a line's "safe setting" on a new skill, check what it does to that skill's reachable range and to the term that makes the spikes, prefer a gain change inside the band the policy was randomized over, verify it in simulation, and re-decide when the contract changes.
Symptom
After the violent first real get-up (2026-08-09), the recovery policy needed a gentler setting for its next hardware test, and the walking and omni lines' standard derating - deploying at power-scale 0.8 - was the obvious candidate.
Context
The V0 recovery contract maps actions to absolute targets over the full joint range: a = +/-1 lands exactly on the URDF limits, and standing puts the knee at the clip. Candidates were compared on R3.1 in MuJoCo (5 categories x 10 seeds) on 2026-08-10 before any hardware time was spent.
Change
A new gain profile, rl_kp090 (kp x 0.9, kd unchanged), recorded in robot.yaml as the recovery hardware-test setting, with power-scale 0.8 explicitly banned for recovery.
Outcome
kp x 0.9: 48/50 got up; median torque demand on hip_pitch/knee fell from 120-125% to 100-104% of the deployment limit; leg-leg contact frames 2,152 -> 1,095; the change sits inside the +/-10% kp randomization the policy trained with. power-scale 0.8: 4/50 could not get up, because under the full-range contract it removes the ends of the travel (the deep squat's tucked legs, the straight standing knee), and the torque spikes (kp x error) did not fall at all. When the line moved to the beta-anchored contract, rl_kp090 was declared a V0-era choice that does not fit (beta is calibrated at kp 30) and deployment returned to rl_default; the deploy switch applies power scaling to the walking side only.
Mechanism
A power scale multiplies the action, which under an absolute full-range mapping shrinks the reachable workspace instead of softening the actuator; the spikes come from the proportional term on large errors, which only a gain change reduces - and a gain change inside the trained randomization band stays in distribution.
Conflicts
The undated operator runbook still carries an R3.1 "B comparison" command at power-scale 0.8 beside the rl_default baseline; the sources do not say whether it was written before the ban or was ever run.
Applies when
- reusing a power, torque or action scale from one skill on another
- a policy whose actions map to absolute targets over the full joint range
- choosing a gentler setting for a first or second hardware trial
“kp×0.9 / kd 不动 —— recovery_r3_1 成功 48/50, τ 需求中位 hip_pitch/knee 120~125% -> 100~104% 部署限, 腿-腿接触 2152 -> 1095 帧; ±10% 在训练 kp DR 带内. ⚠️ power-scale 0.8 对 recovery **禁用**: 全 ROM 契约下 0.8 砍的是行程 端点 (深蹲收腿/站直够不到), 实测 4/50 起不来, 且尖峰 (kp·err) 一点不降 —— 它是 walk/omni 的安全档, 不是 recovery 的.”
git:Lucen-recovery@origin/recovery:robot.yaml § gain_profiles 注释: recovery 真机测试安全档 (2026-08-10) / rl_kp090 Prove an armless get-up exists before training it - a connected static domain, 25% torque on the cheapest path, an 8 mm hand-over gap, a static roll-over - and write down what each scan cannot represent
get-up-feasibility-accounts-before-trainingBefore training a get-up or any multi-contact skill, compute the quasi-static accounts - connectivity of the static domain, torque along the cheapest path, hand-over gaps, COM shift available for rolling - and state which configurations each scan cannot represent; when a policy gets stuck in one of those, extend the scan before blaming the reward.
Symptom
A torso-and-legs robot has no arms to push off the ground; whether it can get up from the floor at all was unknown when the line opened.
Context
recovery_feasibility.py ran three accounts before any training (the run line's "hard accounts first" discipline): a sagittal quasi-static scan (0.05 rad grid, 44,520 configurations, MuJoCo FK, flat-foot assumption). (1) The static standing domain (COM over the feet, torques in limit) has 29,586 cells, flood-fill connected with no islands, from a 0.097 m deepest squat to the 0.384 m stand. (2) The minimum-torque path peaks at 25% of the limits (knee 2.9/12, ankle 3.7/17 N*m) - a 4x margin. (3) All 508 ground-contact configurations have the contact behind the COM; the smallest gap to pure foot support is 8 mm. A roll-over account: swinging both straight legs to one side shifts the COM 96 mm against a 62 mm torso half-width - 1.6x, so rolling needs no momentum. Three conclusions were written down for later attribution: the legs are 80% of the mass (swinging them moves the COM), prone has no flat-foot hand-over face (merge into a supine/side sit first), and supine needs no sit-up (hip flexion is limited to 75 deg).
Change
The accounts gated opening the line and were cited in every later argument about what the robot can physically do.
Outcome
They held where they applied: in V1.0 every fall category was righted under a hard rate limit, which the spec records as the quasi-static roll-over account verified by training, and the 25% torque path was the basis for pursuing a slow get-up. They also misled once: account (3) is sagittal, and on 08-09 the spec corrected its scope - it cannot represent the splayed W-sit where the policy actually stalled. A follow-up prone hip-ROM scan (471,625 cells) found 3,912 two-foot-contact cells and none with both soles within 25 deg of level (best 40.2 deg): a flat-footed push-up from prone is infeasible on this robot, so the fix became where the feet go after sitting up.
Mechanism
A get-up needs a connected path through statically feasible configurations and enough torque along it; quasi-static accounts bound both cheaply, and momentum can only make the real problem easier. A reduced-dimensional scan, though, only speaks for the configurations it can express.
Conflicts
In R0.1-R0.2 the spec read account (3)'s "prone has no front hand-over" as "prone lacks the roll-over skill"; R0.3's confusion matrix showed prone had righted its torso 159/159, and the spec then restricted account (3) to the sagittal configurations it models.
Applies when
- opening a get-up, recovery or climbing skill on a new robot
- a robot lacks arms or other obvious contact options
- a policy stalls in a configuration a feasibility scan never modelled
“本机 **torso + legs、无手臂可撑地**,开训前先证明存在不依赖手臂的物理解 … 双腿同侧直腿摆最大横移 **96 mm = 1.6×** … **静态摆腿即可翻身,无需动量**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §1 机械可行性判决(2026-08-09,recovery_feasibility.py 三笔账) Pre-register the ladder's risks and how each future result will be read - before training
preregister-risks-and-fork-readingsBefore a training ladder or risky rung, write the risks, the stop rules, and how every plausible outcome will be interpreted - then do not edit them after seeing results.
Symptom
Without pre-registration, ladder results get rationalized after the fact; the team had already seen post-hoc reads go wrong and adopted written pre-commitment.
Context
The C-ladder execution sheet opens with three numbered pre-registered risks: (1) S2 plant robustness will not carry into omni - a full S2 redo is budgeted from the start; (2) the s1e recipe has a collapse valley at iter 1500+ (recorded twice), so every rung runs a watch_ckpt --every 100 smoke loop with stop-on-degradation; (3) the root's OOD survival edge may partly be "it is slower / commits less" - with the reading fixed in advance: if C1 training raises tracking while survival drops, the edge was bought with slowness, and root selection reopens. Later rungs went further, pre-registering a full result-to-conclusion table for the A/B arms ("预注册读法（事后不改）") and even pre-registering the author's own doubt that a level would fail and what its failure would prove.
Change
Standing practice: before each rung, write down (a) known risks with their mitigations, (b) the interpretation of each possible outcome, (c) stop criteria - all frozen before the run starts ("开训前写死，事后不许改").
Outcome
When arm-A/arm-B and redo results arrived, conclusions were read off the pre-registered table instead of argued; a predicted-likely-FAIL level (C4-redo3) was still run because its pre-registered value was eliminating the regularization hypothesis - which it did.
Mechanism
Pre-commitment converts each training run into a decisive experiment: outcomes falsify or confirm named hypotheses instead of being absorbed into a story; it also makes negative results valuable (a FAIL that eliminates a hypothesis advances the search).
Applies when
- launching a multi-rung training ladder
- running an A/B fork whose outcome will drive a fork/root decision
- a rung is expected to fail but is run for its diagnostic value
“三条预注册风险 … 若 C1 训后跟踪提上去而存活掉下来，说明这条优势是速度买的、不是通用性 → 那时重开选根。”
train/C_LADDER_RUN.md § 0. 三条预注册风险 A training-side rate limit anchored on the last commanded target is an integrator inside the balance loop - two unrelated lineages converged to the same 34-43% re-fall rate, a soft penalty could not fix it, and the bandwidth arithmetic said safety and standing could not coexist
slew-anchor-is-an-integratorIf you constrain actions in training, anchor the constraint on the measured state, not on the previous command - a limiter with memory adds lag inside the balance loop; and when two different lineages converge to the same failure rate, treat the cause as structural and stop adding soft penalties.
Symptom
With the rate limit moved into training (V1), policies either could not stand up or stood up and kept falling again: the stand oscillated, fell and climbed back, 34-43% of the time.
Context
V1.0-B (user decision: explore new postures from scratch, hard constraint in training): target <- prev + clip(target - prev, +/-S*dt) at the TIGHT rates, anchored on the last issued target like the bridge. Four runs: v1_0 from scratch 0.8% - every category righted under the limit, then knelt (the limit also damped the exploration that had escaped the seated basin in V0); v1_0c continued from R3.1 99.8% get-up but 36-43% re-fall and knee jitter 0.604 rad/s; v1_0p with a pull-assist curriculum (56 N -> 0, fully withdrawn) 39.1% unassisted, re-fall 34-42%; v1_0w with a windup-gap penalty 25.4%, re-fall 18-43%. Removing the limiter from v1_0p gave 0.0%: the policy had co-adapted with it.
Change
The rate-limit route was declared dead after four runs and the line was re-rooted on a beta-anchored action space (V2, user approval required).
Outcome
V2.0's first acceptance at full authority already showed re-falls of 0-1% (V1: 34-43%), the structural bet paying off before any tuning.
Mechanism
Anchored on the previous command, a saturated policy becomes a rate controller - one more integrator in the loop - and active balance through that lag oscillates, while a kneeling sit needs no active control and is stable. The arithmetic: kp 30 needs 0.4 rad of error for 12 N*m; at 4 rad/s that takes 0.1 s, half the pendulum time constant sqrt(0.38/9.8) ~ 0.2 s; keeping standing bandwidth needs S of at least ~8 rad/s, within 20% of the 10 rad/s velocity limit - no bound at all. Anchoring on the measured angle (q + beta*a) makes the full kp*beta authority available in one step, with no memory, and caps the impact at the same time.
Applies when
- adding rate limits, slew limits or target filters to a policy's action path
- a policy stands but oscillates and re-falls after a constraint was added
- different lineages or curricula land on the same failure signature
“v1_0p(拉力课程):会站(prone 79.9%),再摔 34~42% —— **两条完全不同 血统、不同学习路径,收敛到同一失败率**。 … 动作饱和时它退化为 速率控制 = 环内多一个积分器;主动站姿平衡穿过该滞后必振荡(v1_0 的跪坐不需 主动控制,所以稳)。对照:**HoST 的 β 锚在当前实测 q,无记忆、无积分器**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §31 结构病定案:slew 的目标锚 = 控制环里的积分器 Before training a one-leg stand, the accounts and a probe showed the default gains could not hold it at all - kp 20 needs 0.39 rad of error to carry the static roll moment, more than the whole adduction range - so per-joint gains came first, and thermal limits set the session length
single-support-gain-authority-probeBefore training a posture that loads one joint statically, compute the steady tracking error load/kp and the series stiffness against m*g*h, and prove with a simple hand-written controller that the posture can be held under the deployment gains - change the gains first if it cannot; then size session length from the thermal account.
Symptom
The one-leg line (standing on one foot, the other folded back, no hopping) had to decide whether the existing gain profile could hold single support before any reward was designed.
Context
Hardware accounts (9.792 kg, COM 0.234 m high, 170 x 80 mm feet, legs 80% of the mass): moving the COM over one foot needs 107 mm of shift and the 20 deg hip-roll adduction range gives 131 mm - geometrically enough. The static frontal moment is 7.8-9 N*m, within RS02's 17 N*m - torque is enough. But at kp 20 carrying 7.8 N*m needs 0.39 rad of tracking error, more than the entire adduction range, and the real robot had already shown it: commanded +0.17, actual -0.04 (0.21 rad droop) under load, 0.0008 rad hanging - load, not the motor. A probe (probe_oneleg.py) then showed open-loop PD cannot hold single support on physics grounds, so the criterion became "an equilibrium exists and a hand-written 4-gain COM feedback can hold it": single-support roll stiffness is hip and ankle in series and must exceed m*g*h_com = 22.5 N*m/rad; ankle kp 12 in series with hip kp 80 gives only 10.4 (open loop 16/16 fell), ankle 60 with hip 80 gives 34.3 (52% margin).
Change
A per-joint gain profile (rl_oneleg: hip_roll kp 80, ankle_roll kp 60, the rest as rl_default) - which needed per-joint gain support in robot.yaml, the bridge, deploy and the trainer's actuator groups - decided before training. Thermal account: single support makes hip_roll the dominant heat load (about 7.8 N*m against a 7 N*m continuous rating), so acceptance and demos run in segments of at most 60 s with a temperature check.
Outcome
Under rl_oneleg the hand-written feedback held six cells cleanly for 6 s (hip_roll steady torque 2.1-3.4 N*m, half the thermal budget); under rl_default the same feedback on the same cells fell 0/4. The trained V0 policy then passed its 40-cell acceptance.
Mechanism
With PD position control, the steady error needed to carry a static load is load/kp; when that error exceeds the joint's range the posture is unreachable whatever the policy does, and series compliance between joints lowers the effective stiffness below the gravity stiffness that single support demands.
Applies when
- single-support, crouched or one-arm-load postures on PD actuators
- a joint "droops" under load on hardware but tracks well when hanging
- deciding whether a new skill needs its own gain profile
“但 kp=20 时撑住 7.8 N·m 需要 **0.39 rad 跟踪误差 > 整个内收行程**。真机已实测: 命令 +0.17 实际 −0.04（droop 0.21 rad），悬挂时 0.0008 rad——是负载不是电机。 … 单支撑滚转是 hip/ankle **串联**刚度，必须 > m·g·h_com = 22.5 N·m/rad；ankle kp12 串 hip80 只有 10.4（开环 16/16 全摔），60 串 80 = 34.3（裕 52%） … **rl_default 同反馈同格 0/4 全摔**（增益档必要性对照）”
git:Lucen V2@origin/oneleg-line:train/ONELEG_V0_SPEC.md § §1-1 单脚站: 几何可行，卡点是 hip_roll 增益权限 / §2 A 线增益 / §5 probe 定谳 Record-high training reward hid a fully-failing DR subgroup - aggregate metrics average over draws, gates must test per condition
aggregate-metrics-mask-subgroup-failureNever gate on metrics aggregated across DR draws: evaluate at fixed representative conditions (especially the deployment-critical stratum), and if a difficulty axis matters, ramp it on measured per-stratum success rather than sampling the full range from iteration zero.
Symptom
omni_s1e trained under constant-wide latency DR (0, 0.06 s) posted the lineage's highest-ever Isaac reward (129) - while the --delay 2 smoke evaluation showed 3/3 falls from iter 1500 onward, persisting to early stop; the usable checkpoint window shrank to iters 500-1000.
Context
Diagnosis written plainly: "聚合奖励掩盖重延迟尾部子群体失败" - the aggregated reward averages over latency draws, so the majority of light-delay environments can mask the total failure of the heavy-delay tail. The remedy for the training side was a survival-gated ratchet curriculum (survival_gated_latency): the sampling cap starts at 0.02 s and rises +0.01 only when a 4096-reset window's survival (time_out share) reaches >=90%, capped at 0.06, ratchet up-only - "增益与延迟耐受一起长,不升到策略撑不住的 地方" (gain and delay tolerance grow together; never raise past what the policy can hold). The detection side was already in place from the noise-crutch episode: the per-condition smoke curve, not the training reward, is the health readout.
Change
Latency exposure made curriculum-gated on measured subgroup survival instead of uniform-from-zero; per-condition (--delay 2) smoke evaluation kept as the authoritative curve; watcher scoring adjusted (survival weighted 3x) so recovery during hard phases is not early-stopped away.
Outcome
The failure mode was caught by the smoke curve within one generation; the follow-up redesign (deterministic staged latency) superseded the ratchet, but the aggregate-masking lesson held through both.
Mechanism
Expected-return training weights each DR draw by probability, so a subgroup can contribute bounded loss while being catastrophically failed; any scalar averaged over the randomization cannot distinguish "uniformly decent" from "great on easy draws, dead on hard ones". Only conditioning the evaluation on the stratum reveals the split, and curricula should raise difficulty on measured stratum success, not on schedule.
Applies when
- training reward hits records while a fixed-condition eval degrades
- wide DR on an axis where deployment sits at one known value
- designing curricula for difficulty axes (delay, push, terrain)
“常量 latency DR (0,0.06) 从零训被证伪——Isaac reward 129 历代最高,但 --delay 2 冒烟 iter1500 起 3/3 全摔持续到早停(聚合奖励掩盖重延迟尾部子群体失败,可用窗口只剩 500/1000)。… 采样上限 0.02 起步 … ≥90% 才 +0.01s,0.06 封顶,棘轮只升不降。”
train/OMNI_V0_SPEC.md § 3. S1.5(s1e 训练塌方复盘) Training the final recipe from scratch in one run - every mechanism the lineage had accumulated - produced 0% and a seated robot; the order in which the lineage acquired those mechanisms was part of why it worked
curriculum-history-is-part-of-the-productA recipe that ends a lineage is not a recipe for a from-scratch run: consolidate it as ordered curriculum phases matching how the lineage acquired its mechanisms, check that every curriculum criterion is reachable from the starting policy, and read the run's raw term values, not the total reward, before calling it green.
Symptom
V3.0 trained the lineage's whole final recipe from scratch in one 9,000 iteration run - full beta curriculum, prone-conditioned pull assist, friction DR, the 3 s zero gate on standing income, flat_feet - testing the proposition "the product is defined by its configuration, not by its training history". Training looked all green (reward 26.33, episode length 500, 100% time-outs).
Context
Read in raw units against v2_6c at the same weights, the green was a seated equilibrium: base_height 0.377 vs 0.640, stand_pose 0.205 vs 0.516, flat_feet 0.0000 (zero because it sits outside its height gate, not because the feet were flat). Acceptance: 0.0% in Isaac at the deployed authority, 0.0% on every MuJoCo friction level, and still 0.0% at the training-time authority (100% seated at 0.222 m, upright and still).
Change
The full stdout (316k lines) was read: the beta curriculum's criterion (standing share over 0.35) was met zero times, so beta never left the wide setting and the policy had no experience at the deployed authority; the pull curriculum was stuck on the same criterion. The lineage had escaped the seated basin with immediate income (V2.0-V2.2) and only then added the zero gate to cure rushing (V2.5); from scratch, the zero gate removed the early "stand fast, earn more" gradient needed to escape. Verdict "the curriculum history is part of the product", limited to n = 1. v2_6c stayed the product.
Outcome
V3.1 kept the order as explicit phases: P1 from scratch with immediate income (zero gate off) until the curricula advance, P2 adding the zero gate. P1b/P1c escaped the seated basin and passed; P2 was later judged net negative and dropped (time-gate-vs-wide-stance-retire-the-fix).
Mechanism
Mechanisms that refine a competent policy (time gates, tight authority) can delete the gradient a naive policy needs, and a curriculum whose advancement criterion the naive policy never meets freezes at its first level.
Conflicts
The spec limits the falsification to "this recipe + this curriculum criterion" (n = 1, no seed sweep, no criterion tuning). V3.1's phased run succeeding is consistent with the ordering reading but changed other terms too.
Applies when
- consolidating a long lineage of continuation fixes into one clean recipe
- a from-scratch run with all mechanisms enabled plateaus early
- curriculum state is not logged or never advances
“命题:产物由配置定义,而非训练史定义。 … 训练 log(完整 stdout 316k 行)`[beta_anchor]` 仅初始 1 行,**达标 0 次** … V2.0~V2.2 靠**即时计酬**爬出坐姿盆地(§33),站立巩固后 V2.5 才装归零门 治"过快"(§39)。**课程史是产品的一部分。**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §42 V3.0 判决(2026-08-11):从零单 run 全机制 FAIL 于坐姿盆地 The real robot's right-leg kicking was over-trained-delay times loop gain - irreducible pipeline latency is plant, model it fully from day one
pipeline-latency-is-plant-not-drMeasure the end-to-end action pipeline delay and build it into the nominal plant and every acceptance gate from day one; treat power/scale deratings that "fix" oscillation as gain-reduction crutches flagging an unmodeled delay, and expect higher-feedback-gain policies to be MORE delay-fragile.
Symptom
On hardware, s1c/s1d at action scale 1.0 always kicked wildly with the right leg (s1c only ran as SOTA at power 0.8; s1d only at 0.7) - while sim showed nothing under default evaluation.
Context
Sim reproduced the incident item by item once the real pipeline delay was injected: s1d@1.0 with --delay 1 fell at 10.2 s, --delay 2 at 5.2 s; s1c@1.0 stressed (r_hip_roll saturation 5 -> 16%; "右脚" = the policy's chirality makes the right leg its high-gain leg); and the combos that worked on hardware (s1c@0.8+delay2, s1d@0.7+delay2) all survived in sim. Mechanism: the real pipeline is ~1-2 ticks (BusWorker next-cycle pickup + CAN round trip) but S1.1 trained only to 1 tick - "超训延迟 × 全环路增益 = 振荡;衰减 = 压环路增益换稳定" (delay beyond training x full loop gain = oscillation; the power derating had been buying stability by compressing loop gain). s1d was MORE fragile than s1c because its yaw 3-layer stack had learned higher feedback gain - higher gain, lower delay tolerance. Three changes: latency DR widened to cover reality; acceptance gates and smoke runs moved permanently to --delay 2 ("门必须在真机条件下预测 真机"); and the doctrine written twice-paid: "不可约的管线属性(延迟、 限速)不是'随机化选项',是 plant 本体,第一天就该全额建模" - S1's nominal-then-robust staging falsified by hardware for the second time. The later s1e hardware run at power 1.0 (no kicking, normal force) closed the loop: "0.8 = 旧代拐杖" - the derating had been a crutch for the under-modeled delay, not a real requirement.
Change
Latency modeled as plant from day one of any lineage (measured 1-2 ticks covered, bridge-layer rate limits likewise modeled by default); every gate and smoke evaluation issued under --delay 2.
Outcome
Kicking reproduced, explained, and eliminated in the s1e generation at full scale and full power; the deploy-side crutches (0.7/0.8) retired for the new lineage.
Mechanism
Feedback oscillation onset is a product of loop gain and phase lag; a policy trained below the real delay learns gains that sit past the real stability margin, and any output derating masks it by scaling gain down. Since pipeline delay is deterministic hardware property - not an uncertainty - it belongs in the nominal plant, and every evaluation must include it or the gate predicts a robot that does not exist.
Applies when
- hardware oscillation/kicking that sim only reproduces with added delay
- a policy only runs on hardware at reduced power/scale
- defining what belongs in the nominal plant vs the DR list
“真实链路延迟 ~1~2 拍 … S1.1 只训到 1 拍——超训延迟 × 全环路增益 = 振荡;衰减 = 压环路增益换稳定。s1d 比 s1c 更脆 = yaw 三层栈学出更高反馈增益,增益越高延迟容忍越低。… 教训入账:S1「先标称后鲁棒」第二次被真机证伪——不可约的管线属性(延迟、限速)不是"随机化选项",是 plant 本体,第一天就该全额建模。”
train/OMNI_V0_SPEC.md § 3. S1.4(真机右脚乱踢事故强制) A proposal in the runbook - torque and action limits as versioned safety tiers (classroom / research / expert) written to motor RAM and read back, separate from the reward's effort penalty - recorded as a proposal, its implementation unrecorded
safety-limits-are-a-layer-not-a-rewardKeep hardware limits as an explicit, versioned safety layer (tiers written and read back at start, the persisted default the safest one) and the effort penalty as a behaviour layer; when a skill needs more torque, change tier deliberately rather than trading one layer against the other.
Symptom
Running and jumping need more torque than the deployed limits allow, and the temptation is to trade the training-side effort penalty against the hardware limit, or to hand a new user a robot "tuned however the last person left it".
Context
A message pasted into the operator runbook (undated, citing Berkeley's practice of storing the full motor configuration as JSON with write and read-back scripts) proposes: configuration is a versioned artifact, not a verbal agreement; three safety tiers in robot.yaml beside the gain and policy profiles - classroom (RS06 limited to 10 N*m, lateral joints clamped: "however bad the policy, it only moves awkwardly"), research (14 N*m, clamps at twice the measured need, the default) and expert (the 36 N*m rating, joint limits only, requiring an explicit flag); deploy writes the tier to motor RAM at start and reads it back, while the stored copy stays classroom so a power cut returns to the safest state. It frames limits as the safety layer and the effort penalty as the behaviour layer - more torque for running means switching tier, not weakening the penalty.
Change
None recorded: the message ends by asking which to do first, a rollback or the tiers.
Outcome
The sources do not record the tiers being implemented; the deployed limits stayed at 12/17/11 N*m through the recovery and one-leg lines (the one-leg spec treats raising the RS06 limit as a separate, unapproved hardware decision). Related and recorded elsewhere: torque limits were written to RAM only in a scripted, self-reversing field experiment.
Mechanism
Hardware limits bound the damage any policy can do; reward terms shape what a policy prefers. Mixing them either weakens safety to buy behaviour or distorts behaviour to buy safety.
Applies when
- a new skill needs more torque than the deployed limits
- robots are handed to students or new users
- motor configuration lives in people's heads or in the firmware only
“配置是版本化的产物,不是口头约定。 … deploy_policy 启动时按档写进电机 RAM 并读回校验(落盘的那份永远保持 classroom,断电自动回到最安全状态)。 … 限幅是安全层,dof_torques_l2 是行为塑造层,它们在不同的层,不冲突。跑步要更大力矩就换档,而不是去动训练里的省力惩罚。”
RL系统/FOLLOW THIS copy 2.md § 面向 developer / 教育机构该怎么做 (pasted proposal, undated) The latency DR range must cover the measured deployment pipeline - 0-20 ms could not even reach the real 1-2 control steps
latency-dr-covers-measured-pipelineMeasure end-to-end action latency in control steps on your own stack (including cross-process queue boundaries), set the DR range to cover it with margin, and never import a delay count without its control frequency.
Symptom
Action latency was randomized over 0-20 ms (0-1 control step at 50 Hz), but the measured deployment path is 1-2 steps: the deploy process writes the target, an independently running BusWorker picks it up on its NEXT cycle, plus CAN round-trip - the training range could not cover the robot's actual latency at all.
Context
Fix: widen action_latency_s to 0-0.06 (0-3 steps). The external reference's "uniform 6 steps" was explicitly NOT copied - that number depends on his unknown control frequency; locally, a sweep at 0/1/2/3 steps showed walk_v5 survives all with insensitive metrics, so 6 steps "在我们这里没有依据" (has no local basis). The range was set from the measured pipeline with margin, not from a foreign constant.
Change
action_latency_s (0, 0.02) -> (0, 0.06), justified by pipeline analysis (writer/worker cycle boundary + bus time) and bounded by the local latency sweep.
Outcome
The DR band now brackets the true deployment latency; the policy trains against the delay it will actually face instead of a fictional sub-step world.
Mechanism
Latency DR only immunizes against delays inside its support; a range below the physical pipeline guarantees an untrained distribution shift at deployment. The correct range comes from tracing the pipeline's worst case (queueing boundaries + transport), and foreign step-counts are meaningless without the control rate they were measured at.
Applies when
- setting or auditing action-delay randomization
- deployment uses a separate bus/worker process from the policy loop
- importing delay-modeling numbers from other projects
“现行 0~20 ms = 0~1 个 50Hz 控制步, 而实测部署链路是 1~2 步(deploy 写 STATE.target 后, 独立跑的 BusWorker 下一轮才取走下发, 再加 CAN 往返)——现在的区间覆盖不到真机的实际延迟。… 不照抄参考来源的"统一 6 步": 那取决于他的控制频率(未知), 而我们扫过 0/1/2/3 步 … 6 步在我们这里没有依据。”
train/WALK_V7_SPEC.md § ⑤ action_latency_s 0~0.02 → 0~0.06 A 2-degree joint-zero calibration fix moved the whole runnable envelope - re-test old "cannot run" verdicts after recalibration
zero-offset-calibration-shifts-envelopeDate every hardware verdict with the calibration state; after any zero/mount recalibration, re-test previously condemned policy-power combinations and previously "unexplainable" posture offsets before attributing either to training or model.
Symptom
s1d was on record as "only runs at power 0.7" (kicked wildly at 0.8); after a calibration pass, the same policy ran 12 s at 0.8 with no kicking at all.
Context
The calibration had fixed a 2.08 deg zero offset on r_hip_roll - exactly the constant error source on the dominant joint of the kicking oscillation loop ("恰是乱踢振荡环主导关节的常值误差源"). The three-generation post-calibration hardware sweep also closed a second case: the robot's mysterious "backward lean" disappeared after calibration, and the sim-real posture difference collapsed from opposite-sign 5+ deg to same-sign ~2 deg ("后仰案实质了结") - the lean had been a sensing/zero artifact, not a mass-model error. Booked consequence: if the s1d recovery re-verifies, "真机可跑档整体 上移" - every policy's runnable power envelope shifts up, and downstream lineages' hardware expectations get revised.
Change
Joint-zero and mount calibration promoted from setup chore to a variable that dates hardware verdicts: verdicts about which power/scale levels a policy can run are conditioned on the calibration state they were measured under.
Outcome
One policy rehabilitated at a higher power level; one standing sim-real posture discrepancy closed without touching model or training; a pending re-verification booked rather than asserted.
Mechanism
A constant joint-zero error acts as a persistent disturbance injected at the feedback loop's most-loaded joint; near an oscillation threshold, removing a 2-degree bias is the difference between a stable and an unstable loop. Since the error is additive and machine-side, it shifts every policy's stability envelope simultaneously - which is why verdicts must carry their calibration date.
Conflicts
The s1d rehabilitation awaited one confirming re-run at the time of writing ("待复核一跑坐实") - the offset-as-cause reading is the head suspect, not a closed verdict.
Applies when
- a policy oscillates at a power level others tolerate
- sim and real disagree on a constant posture offset
- deciding whether to re-test old hardware verdicts after maintenance/calibration
“发现①:s1d@0.8 能跑了(旧账「只有 0.7 能跑」)——12s 无乱踢。头号嫌疑 = 标定修正:r_hip_roll offset 修 2.08°,恰是乱踢振荡环主导关节的常值误差源。… 发现②:「后仰」标定后消失 … sim-real 姿态差从反号 5°+ 收敛到同号 2°,后仰案实质了结。”
train/README.md § 真机 @0.8 三代横评(2026-08-07 标定后)