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
115 cards matching “fine-tune-reward-change-falsified”.
Fine-tuning through a reward-table change was falsified (scatter, half-recover, collapse) - continuation training is legal only with the reward frozen
fine-tune-reward-change-falsifiedNever fine-tune through a reward-table change - retrain from zero; reserve checkpoint continuation for frozen-reward plant/DR widening, reset noise_std when branching, and watch for the scatter/half-recover/collapse signature as the abort trigger.
Symptom
The s1c A/B experiment: arm B fine-tuned from an existing checkpoint under the revised reward (same contract, same network, changed reward + small DR) and failed with a characteristic signature - scatter, half-recover, fall back ("打散→半恢复→摔回"); arm A trained from zero under the same config won decisively (full shaping lifted swing to 21.6 mm within 500 iters; shipped at 5500).
Context
Verdict recorded: "从零 + 强塑形是本机唯一验证过的发育路径" (from-zero plus strong shaping is this machine's only validated development path). The signature became a standing stop criterion in every later rung that touched a reward ("s1c B 臂签名,出现即停"). Crucially the boundary of the law was drawn explicitly when S2 continuation training was proposed: "当年证伪的是「奖励表中途改版的 fine-tune」… S2 续训奖励表全程冻结、只逐级加宽 plant/DR,属另一类" - continuing a checkpoint with the reward FROZEN while widening plant/DR one rung at a time is a different class and was allowed (and then worked, powering the whole S2/C lineage) - with the honest fallback that if frozen-reward continuation ever collapses, that rung retrains from zero and the doctrine gets re-examined with data. Fine-tune arms also need mechanical care: reset the checkpoint's collapsed noise_std (terminal 0.033 "会杀死探索") and account for iteration counters re-zeroing (curriculum gates fire immediately).
Change
Reward changes and lineage continuation permanently separated: reward revisions -> from-zero retrain; plant/DR widening -> frozen-reward continuation with per-rung gates; the B-arm signature promoted to a universal tripwire.
Outcome
No later reward revision was attempted by fine-tune; frozen-reward continuation carried S2 (PD/COM/friction rungs) and the C command ladder successfully from the s1e root.
Mechanism
A trained policy sits in an optimum of its reward's geometry; changing the reward moves the optimum but leaves the policy's exploration noise near-zero and its value function calibrated to the old returns - it disassembles the old solution faster than it can assemble the new one. Widening DR under a frozen reward instead keeps the optimum's identity and asks only for local robustification.
Applies when
- proposing to fine-tune an existing policy under a revised reward
- planning a robustification ladder from a validated checkpoint
- a continued run scatters then partially recovers then collapses
“B 臂 fine-tune 证伪(打散→半恢复→摔回——从零 + 强塑形是本机唯一验证过的发育路径)。… 当年证伪的是「奖励表中途改版的 fine-tune」(B 臂,塑形突变致终盘摔回);S2 续训奖励表全程冻结、只逐级加宽 plant/DR,属另一类;若 s2_lag1 续训本身塌方,回退方案 = 该级从零重训,续训教义再议(拿数据说话)。”
train/OMNI_V0_SPEC.md § 3. S1.3 / 4. 与 s1c fine-tune 证伪的关系 Cutting swing amplitude 40% raised yaw-momentum demand 53% - the falsified fix is recorded so nobody walks that road again
amplitude-cut-falsified-yaw-fixTest gait fixes against the quantity the ground must actually supply (torque/force rates vs friction ceilings), not against kinematic proxies; record falsified fixes with their mechanism so the search space shrinks permanently.
Symptom
Support-foot yaw slip stayed at ~223 deg (vs 225 deg) after walk_v6 cut joint swing amplitudes by ~40% (hip_pitch 20.6->10.4 deg, knee 31.3->18.8 deg) - the change built on the theory "smaller swing = less yaw momentum to dump into the ground".
Context
Direct measurement inverted the theory: yaw-momentum amplitude ROSE 16% (+/-0.1000 -> +/-0.1159) and its rate of change rose 53% (1.86 -> 2.84 N*m demanded from the ground), pinned exactly at the foot's supply ceiling (2.8-3.3 N*m at mu 0.6-0.7) - so slip could not drop. The extra demand lives in higher harmonics: v6's crisper foot placement (better clearance 34 mm, lower landing force) shortens the momentum- exchange window, concentrating the same exchange into less time. The verdict was written as a closed road: "下一轮不要再走这个方向". An honest residue was also booked: WHY amplitude down but momentum up 16% remained unresolved, with the named next step (per-rigid-body decomposition of H_z, since per-joint RMS sensitivity ignores phase correlations).
Change
The "reduce amplitude to reduce yaw momentum" lever was removed from the planning space; future yaw-slip work redirected toward the supply side (friction) and momentum-rate mechanics.
Outcome
Slip unchanged (225 -> 223 deg); the falsification and its mechanism became a permanent constraint on the fix search space.
Mechanism
Ground yaw torque demand scales with the rate of change of angular momentum, not its amplitude; kinematic amplitude cuts that also sharpen contact timing can raise dH/dt while lowering range. When demand exceeds the friction-limited supply ceiling, slip is set by the ceiling, so demand-side changes below the ceiling do nothing visible.
Applies when
- attacking foot slip or yaw drift via gait shape changes
- a fix targets an amplitude while the constraint is a rate
- documenting a failed intervention after a version comparison
“walk_v6 | ±0.1159 (+16%) | 2.50 Hz | 2.84 N·m (+53%) … 脚的供给上限 2.8~3.3 N·m(μ 0.6~0.7)—— v6 正好顶在天花板上, 所以滑移一点没降。… 结论: "减小摆动幅度以降低偏航动量"这条被 v6 证伪 —— 砍 39% 幅度, 需求反升 53%。下一轮不要再走这个方向。”
train/WALK_DIAGNOSIS.md § ② 未生效的机理: 减小摆动幅度反而让偏航需求上升 action_rate weight is the sim2real bandwidth knob - re-tune it whenever a rate limiter is removed
action-rate-weight-vs-bandwidthSet action_rate weight relative to real actuator bandwidth, and re-tune it any time another smoothing/limiting element (filter, slew limiter, gain) changes - reward weights are load-bearing parts of the actuator model.
Symptom
With a low action_rate_l2 weight the policy learns fast actions; the unmodeled part of the actuator response is then excited hardest, and sim2real "直接崩" (collapses outright). With too high a weight, actions become so slow the robot cannot maintain balance.
Context
The reference developer called action_rate_l2 the single most important reward for transfer, with side-by-side video evidence that the high-penalty, slower policy is clearly better on hardware. Lucen context: the team had just removed the SOFT_SPD=1.0 velocity limiter, which had been an implicit actuator-bandwidth constraint - leaving action_rate as the only remaining constraint on action speed.
Change
Decision recorded: after removing SOFT_SPD, re-evaluate the action_rate weight rather than keep the old value, since its effective role changed from "additional smoother" to "sole bandwidth constraint".
Outcome
Logged as a priority follow-up ("重新评估 action_rate 权重 - 拆掉 SOFT_SPD 之后这一项的作用变了"); the failure mode it guards against is training high-frequency actions the real actuators cannot track.
Mechanism
Slower actions stay inside the frequency band where the ideal-PD sim actuator and the real actuator agree; fast actions probe the band where unmodeled delay, inductance, and bandwidth limits dominate, so model error is amplified in exact proportion to action speed. Any removed external rate limit transfers that constraint's entire job onto the action_rate penalty.
Conflicts
The low/high tradeoff evidence is the external developer's report (with video); the Lucen-side entry is a pre-registered risk and decision, not yet an on-robot A/B at the time of writing.
Applies when
- removing or adding an action filter, slew limiter, or low-level speed cap
- real robot shows high-frequency chatter or overheating absent in sim
- tuning smoothness rewards before a hardware deployment
“权重低 → 动作快 → 执行器模型不准的部分被放大，sim2real 直接崩 / 权重高 → 动作慢 → 好迁移，但可能慢到无法维持平衡 … 我们刚拆掉 SOFT_SPD=1.0 的限速器，等于把执行器带宽约束整个移除了。action_rate 惩罚现在是唯一还在约束动作速率的东西，需要重新评估权重”
Experience.md § action_rate_l2 是他认为最关键的 reward (lines 61-70) Gate a new reward term by its command so all old modes score pointwise identical
gate-new-reward-terms-by-commandWhen a reward term must be added mid-lineage, gate it on the condition that defines the new task so every pre-existing situation scores exactly as before - and still watch for value-rescale pathologies inside the new mode.
Symptom
Adding a lateral tracking reward (track_lin_vel_y_exp) ungated would have paid 0-2.0 per step even in modes with cmd_vy = 0 (healthy gait sway of vy ~0.1 already earns 1.28), shifting the whole reward table by a large bias and rescaling the value function - no longer "just adding one mode".
Context
C4 was the C ladder's only true reward surgery. Single-variable discipline required that the change be invisible to every existing mode. The chosen construction: gate_by_cmd=True - the term pays only when |cmd_vy| > 0.02, so for all modes with cmd_vy == 0 the term is pointwise zero, i.e. the reward is pointwise identical to before the change. The same trick appeared earlier in C1: replacing the vy L2 tax with a command-error version that is "对 cmd_vy≡0 逐点同值" (pointwise equal when cmd_vy is 0), explicitly classified as not-a-reward-change.
Change
track_lin_vel_y_exp added with gate_by_cmd=True (weight +2.0, std 0.15); the residual acknowledged honestly - inside the side bucket the values DO change, so the rung still watched the known reward-reshuffle pathology signature (s1c B-arm: scatter -> half-recover -> collapse) as a stop criterion.
Outcome
Old modes provably unaffected (pointwise-equal argument); attribution for any change in old-skill metrics stayed clean through the C4 redo series.
Mechanism
PPO's critic normalizes to the reward scale it sees; an ungated additive term shifts returns in every state and re-scales advantages globally, entangling the new skill with all old ones. Command-gating confines the new term's support to the new mode's state distribution, making "pointwise identical elsewhere" a provable property rather than a hope.
Applies when
- adding a tracking/shaping term for a new command or skill to a lineage that must not regress
- reward change proposed while other skills are still being gated
- reviewing whether a config diff counts as a reward change
“只在 |cmd_vy| > 0.02 时付。不门控的话它对 cmd_vy≡0 的老模式也给 0~2.0 分（健康摇摆 vy≈0.1 → 1.28），等于给整张奖励表加一个大偏置、值函数尺度全变 … 门控后老模式逐点得 0 = 与加项前逐点同值，单变量纪律成立。… 但 side 桶内的值确实变了 —— 这仍是奖励表改版，开级盯 s1c B 臂签名”
train/C_LADDER_RUN.md § 3d. gate_by_cmd=True（重要） Compare the achieved reward to the computed ignore-floor to tell "never learned" from "learned but unprofitable"
ignore-floor-diagnosisFor any skill that trains flat, compute the reward the null policy would earn on that term; achieved==floor means the behavior never paid out (find why: exploration, reward observability, or feasibility) - do not tune weights first.
Symptom
C4 sidewalk failed on both arms; the question was whether the policy had found sidewalk and rejected it as unprofitable, or never found it at all - two diagnoses with opposite fixes.
Context
The theoretical value of the sidewalk tracking term for a policy that completely ignores the command was computable from the command distribution: 0.189. Trained final values landed at 0.187 (arm A) and 0.204 (arm B) - sitting exactly on the ignore-floor - while the honest balance at the optimum actually favored sidewalking (net +0.88/step inside the side bucket). Later the same arithmetic closed the whole saga: for the true reward landscape, doing real sidewalk scored 0.153 vs 0.944 for ignoring - the policy's refusal "是理性最优，不是探索失败" (rational optimum, not exploration failure) under one hypothesis, and under the final measurement-corrected account the policy had "每一次都在 做理性选择" (made the rational choice every time).
Change
Diagnostic rule adopted: compute the ignore-floor for the new term; if the achieved value sits on it, the behavior was never expressed in useful volume (or the reward cannot distinguish it - check both); if the achieved value is above floor but the behavior is absent at deployment, the policy sampled it and priced it out - then the reward balance, not exploration, is the lever.
Outcome
Correctly identified that PPO had not merely under-valued sidewalk; each subsequent hypothesis (waveform sign, regularization cage, exploration form, reward kernel) was tested against this floor arithmetic, which kept the search honest through three reversals.
Mechanism
Every reward term has a computable value under the null behavior; the achieved-vs-floor gap is a one-number audit of whether the optimizer ever monetized the target behavior. It converts "training failed" into one of two mechanistically distinct states with different fixes.
Applies when
- a new skill's tracking reward plateaus early
- deciding between exploration fixes and reward-weight fixes
- post-mortem of a failed curriculum rung
“track_lin_vel_y_exp 训练终值恰好坐在「完全无视指令」的底分上（A 0.187 / B 0.204，理论值 0.189），而终点 balance 明明有利（side 桶内净 +0.88/步）。不是学会了不划算，是根本没学到。”
train/C_LADDER_RUN.md § 3e. 为什么首战 FAIL Before adding a command mode, compute what ignoring it costs - the lazy optimum must lose
reward-cost-of-ignoring-auditPrice the do-nothing policy for every new command or objective: compute reward-per-step for "comply" vs "ignore" from the actual table, and only train once ignoring is decisively unprofitable.
Symptom
A new command mode can be silently unlearnable if the reward table makes "ignore the command entirely" nearly free compared to the tracking reward available elsewhere.
Context
For each C rung the team computed the per-step cost of completely ignoring the new command versus ignoring forward: ignoring vx=0.25 costs 1.264/step; ignoring wz=0.20 costs 0.984/step (78% of forward - gradient sufficient, so C2 was certified "zero reward surgery"); but ignoring vy=0.10 cost only 0.020/step - 50-60x weaker, because vy entered the table only as an L2 tax, not a tracking term.
Change
Rule instituted: a rung may claim "no reward change needed" only after this arithmetic shows the ignore-cost is the same order as forward's. For C4 the audit failed, so track_lin_vel_y_exp (+2.0, std 0.15, same form as vx) was added - raising the ignore-cost at cmd_vy=0.10 from 0.020 to 0.718/step (36x).
Outcome
C1/C2/C3 proceeded with zero reward edits, keeping single-variable attribution clean; C4's needed surgery was identified before training instead of after a failed run.
Mechanism
PPO converges to whatever costs least; if the reward margin for obeying a new command is a rounding error against existing terms, the "ignore" policy is the optimum and no amount of training fixes it. The audit prices the lazy optimum explicitly before spending compute.
Applies when
- adding a command axis or task mode to an existing reward table
- a new skill trains flat while other skills stay healthy
- certifying a rung as "no reward change"
“cmd wz 0.20 → 0.984（coarse .473 + fine .491 + L2 .020）… 对照：忽略 vx=0.25 = 1.264（本级 78%，同量级）；忽略 vy=0.10 = 0.020（弱 50 倍——那才是 C4 必须加 track_lin_vel_y_exp 的原因）。本级不动奖励表。”
train/C_LADDER_RUN.md § 4. 原地转级（C2）② 奖励梯度已验够 Continuing a converged policy on a change that carried no new gradient drifted its transfer from 100/98% to 80/28% over 3,000 iterations while every Isaac gate stayed perfect - scan every checkpoint on the second simulator's friction axis
converged-continuation-is-poisonBefore continuing a converged policy, check that the change creates a live gradient; if it does not, cap the budget at a few hundred iterations, and in every continuation scan each checkpoint on the second simulator's transfer axis (for example low friction) - trainer-side gates can stay perfect while transfer decays.
Symptom
V2.7-A (swap the flat_feet term for a compensated version, continue from v2_6c) finished with the line's best Isaac score (100%) and a MuJoCo transfer collapse: mu 1.0 98 -> 80%, mu 0.4 98 -> 28%; the stance it was meant to widen had not moved.
Context
The new term's calibration run showed a near-zero tax from the start: the policy already satisfied it, so the reward landscape offered nothing new. A checkpoint scan on MuJoCo mu {1.0, 0.4} located the damage: +100 iterations 100/98% (better than the baseline), then 86/54, 60/38, 80/28 - monotonic decay with training length, while entropy and action noise rose (7.77 -> 8.18, 0.588 -> 0.612): drift, not sharpening.
Change
Rule written in: with no new gradient, a continuation budget is short (at most a few hundred iterations) and the MuJoCo transfer axis enters every checkpoint scan. The next rung (V2.7b, a live stance-width gradient) was budgeted at 1,000 iterations with mu {1.0, 0.4} scans every 100 and a stop-on-signal rule.
Outcome
V2.7b kept transfer at the same depth (mu 1.0 98% / mu 0.4 92% at +1,000, where A had already rotted to 86/54) and at +3,000 (100/96%): a live gradient preserved transfer. V2.8 then broke that pattern (mu 0.4 2%): the gradient must also be compatible with the policy's existing form.
Mechanism
On a converged reward landscape PPO keeps updating without a signal to follow, and the random walk is pulled toward whatever the training plant rewards idiosyncratically - invisible in the trainer's own gates.
Conflicts
The drift mechanism is the spec's reading of one decay series plus one contrasting run; V2.8 is recorded as an exception to "live gradient keeps transfer".
Applies when
- fine-tuning a converged policy with a small reward change
- a continuation run's trainer-side metrics improve while real or cross-sim results worsen
- choosing which checkpoint of a continuation to ship
“**checkpoint 扫定死因**(μ1.0/μ0.4):**29500(+100 iter)= 100/98%** (优于基线!)→ 30400 = 86/54 → 31400 = 60/38 → 32398 = 80/28 —— **迁移随续训长度单调衰减**。 … **教训入库:收敛均衡上的长续训是毒药 —— 无新梯度时 续训预算须短(≲数百 iter),且 MuJoCo 迁移轴必须进 checkpoint 扫描。**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §43 结果:V2.7-A 判 FAIL —— 换刀本身无罪,毒在续训预算 The restart reward table lists a reason for every term AND a lesson for every exclusion - absent terms are removed, not zero-weighted
minimal-reward-table-with-provenanceMaintain the reward table as an evidence ledger: every term cites the episode that justifies it, every excluded term cites the episode that convicted it (including the development stage it is valid at), and retired terms are deleted from the config, never left at weight zero.
Symptom
Seven walk generations had accumulated an entangled reward table where nobody could say which term earned its place; the restart needed a table that could be audited line by line.
Context
The minimal table v2 was built under three written principles: "一项管 一件事、结构性反抬脚的项一个不留、塑形只留一套相位逻辑;不在表里的一律 不加" (one term per job; zero structurally-anti-lift terms; exactly one phase-shaping logic; nothing outside the table gets added). Every row carries its provenance (e.g. base_height target = standing height cites the v4 crouch lesson; split x/y tracking cites the merged-exp gradient hole; world-frame yaw cites the v1/v3 body-frame lesson). Every EXCLUSION carries its same-type precedent: feet_landing_vel out because it is poison while the gait is unformed (drag pays 0, lifting pays - the v4-clearance / v8a-B "reverse threshold" family) though it was fine in v7 when the gait already existed - term validity depends on developmental stage; feet_air_time out with its measured non-lever evidence (v6 had it at 2.0 and still lifted 4 mm); and replaced terms are REMOVED from the config ("置 None,不是权重 0 挂着") so audits see truth, not dormant weight.
Change
Reward table rebuilt as ~20 rows each with weight + provenance column; exclusion list maintained alongside with the falsifying episode for each; dormant terms deleted rather than zeroed.
Outcome
Later revisions (S1.2/S1.3) modified the table by citing and updating specific rows' evidence rather than re-arguing the whole design; the table doubled as the lineage's reward-lesson index.
Mechanism
A reward table is a set of standing hypotheses; attaching each row's evidence makes revisions targeted and reversible, and recording why a term is absent prevents the cycle of re-adding known poisons. Deleting vs zero-weighting matters because config audits and DR interactions see the term either way - a zero-weight term is dormant complexity waiting to be flipped on wrongly.
Applies when
- designing a reward table for a restart or new task
- someone proposes re-adding a previously removed term
- auditing which reward rows still earn their place
“原则:一项管一件事、结构性反抬脚的项一个不留、塑形只留一套相位逻辑;不在表里的一律不加 … feet_landing_vel(评审 #4):拖地时代价恒 0、抬脚才收费——与 v4-clearance/v8a-B 同属「反抬脚门槛」家族,步态未成形时是毒;v7④ 加它时步态已存在。… 已从 cfg 移除(置 None),不是权重 0 挂着。”
train/OMNI_V0_SPEC.md § 3. 最小奖励表 v2 / 明确不带 Real robot walked at half the sim clock for two generations - resolved by racing a reward-side and a plant-side evidence line, not by guessing
period-doubling-evidence-raceFor a hardware-only pathology, refuse to guess: pre-register one probe per side of the sim2real boundary (can the reward mechanism change it on hardware? can fitted plant parameters reproduce it in sim?) and let the first positive result direct the next version.
Symptom
The number-one sim2real gap: on hardware v6/v7 stepped at 1.23-1.32 Hz - almost exactly half the 2.50 Hz gait clock they were trained and simulated at; sim never reproduced it, two generations running.
Context
Instead of committing training budget to a guess, v8 pre-registered two mutually controlled evidence lines and kept the clock OUT of the training variables: (a) reward-side - if the v8 saturation fix revives joint_pos_ref (the term that pins the gait to the clock), re-run hardware and see whether frequency returns to 2.5 Hz (hypothesis: v7's frozen actions meant NO reward was pinning the gait to the clock, and the real plant - with armature and friction making high frequencies expensive - slid down to the leg's pendulum natural frequency ~1.1 Hz); (b) plant-side - record suspended joint data (fit_actuator), fit armature/friction, load the fitted values into sim2sim and see whether the 1.25 Hz reproduces IN SIM. Decision rule fixed in advance: "谁先给出阳性结果谁定 v9 的方向 (奖励侧 vs plant 侧)" - whichever line goes positive first sets the next version's direction.
Change
Period-doubling excluded from the v8 change set; both diagnostic lines scheduled in parallel as non-blocking work; frequency reported factually in acceptance with no pass/fail attached ("倍周期是否消失 不设判定,它是 §9 的关键证据").
Outcome
The gap was routed into a decisive-experiment structure rather than a speculative retrain; the plant-side line pointed at exactly the unmodeled armature/friction that were later measured and installed as the plant baseline. Resolution (era-2c full-plant retest): the family had TWO causes - v8's low-speed period-doubling vanished once measured armature+friction were installed (1.30 -> 2.50 Hz, bifurcation-edge machine sensitivity), while v7's stood untouched at 1.20 Hz (saturation-freeze-driven policy property) - both evidence lines paid off, one per case.
Mechanism
A behavior appearing only on hardware has candidate causes on both sides of the sim2real boundary; changing training to fix it tests only one side per expensive cycle. Two cheap parallel probes - one intervening on the reward mechanism, one making sim reproduce the real behavior - localize the cause to a side before any training money is spent, and sim-reproduction of a real pathology is itself the strongest form of plant validation.
Applies when
- a gait pathology appears on hardware but never in any simulator
- deciding whether a sim2real gap is reward-side or plant-side
- tempted to change the gait clock/reward to chase a hardware symptom
“倍周期(真机 1.23~1.32 Hz ≈ 时钟一半,v6/v7 连续两代;sim 从不出现):两条证据线互为对照——(a)… 真机重跑看频率是否回 2.5 Hz(假说:v7 没有任何奖励把步态钉在时钟上,真机 plant 有 armature/摩擦、高频贵,自由滑落到复摆自然频率 ~1.1 Hz);(b)真机吊挂录 fit_actuator.py … 看能否在仿真里复现 1.25 Hz。谁先给出阳性结果谁定 v9 的方向。”
train/WALK_V8_SPEC.md § 9. 平行线 (倍周期) 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 ⚠️ Multiple changes may share one rung only if their symptom spaces are orthogonal - with the ablation order written in advance
orthogonal-batch-with-ablation-orderBatch changes into one rung only when you can name each change's private symptom space in writing; pre-register the ablation order (numeric before structural) and per-change escalation plans, so a mixed outcome decomposes without new decisions.
Symptom
v8 needed four repairs at once (saturation cheating, landing impact, leg narrowing, heading alignment) - strict one-variable laddering would have cost four training cycles for wounds that were all already diagnosed.
Context
The batch was allowed because each change owns a disjoint symptom space, stated explicitly: "A→形态/饱和, B→落地/步态高度, C→腿距/roll 摇摆, D→偏航/转向" - so a single run can attribute each outcome to its change by which symptom moved. For the failure case, the ablation order was pre-registered (A2 -> A1 -> B -> D -> C, "先撤数 值改动" - retract numeric tweaks before structural ones), and every change carried its own escalation/rollback plan (e.g. A insufficient: joint_pos_ref 1.6->2.0 or widen the exp kernel; B overdone - robot afraid to land: v_ok 0.30->0.45; D unstable: rel 0.5->0.25, not back to 0).
Change
Four-change rung executed as one run with per-change symptom ownership, per-change contingency plans, and a pre-registered global ablation order for unattributable regressions.
Outcome
The rung retained single-run attributability without paying 4x training cost; the contingency table meant no failure mode would require improvising an ablation under pressure.
Mechanism
The one-variable rule exists to keep attribution possible, not as an end in itself; attribution survives batching exactly when the changes' observable effects are separable. Orthogonality is a claim that must be argued per pair in advance - and the pre-registered ablation order is the escape hatch for the case the claim fails.
Applies when
- several diagnosed fixes are queued and ladder time is scarce
- deciding between strict laddering and a combined rung
- a combined rung shows a regression no single change explains
“四个改动症状空间基本正交,可单 run 归因:A→形态/饱和,B→落地/步态高度,C→腿距/roll 摇摆,D→偏航/转向。出现无法归因的整体退化时消融顺序 A2→A1→B→D→C(先撤数值改动)。”
train/WALK_V8_SPEC.md § 8. 风险与归因 Changing the gait clock silently flipped a hardwired threshold's meaning - write derived constants as expressions
derived-constants-must-track-their-baseBefore changing any base parameter (clock, control rate, scale), enumerate every constant derived from it and every constant that must NOT change; convert derived literals into expressions of the base so the next change cannot silently flip a term's meaning.
Symptom
Slowing the clock 0.40 -> 0.50 s would have silently inverted the feet_air_time threshold's semantics: the 0.25 s threshold was hardwired, so at ct 0.40 the swing window (~0.20 s) sat below it (constant pressure to lengthen strides), while at ct 0.50 the window (~0.25 s) equals it - the term's meaning flips from "push longer" to "neutral" with no code error anywhere.
Context
The clock change audit walked every dependent quantity: most followed automatically (joint_pos_ref / clearance / contact_number cycle_time params, gait_phase observation, deploy/sim2sim/policy_io, export) - wiring confirmed, zero hand edits; the air_time threshold was the one hardwired constant, fixed by preserving the RATIO: 0.25 -> 0.3125 = 0.625 x ct, with the recommendation to commit it as the expression 0.625*ct "一劳永逸" (solved once and forever). The same audit also listed what must NOT follow the clock (50 Hz control rate, physics dt/decimation, 47-dim contract, action_latency absolute seconds, PD/torque limits) - the change's blast radius stated in both directions.
Change
feet_air_time threshold re-expressed as a fraction of cycle_time; auto-following vs must-not-change lists written into the spec for the clock migration.
Outcome
The clock migration (v10, repeated in v11) carried no silent semantic flips; the expression form removed the trap for every future clock change.
Mechanism
Constants derived from a base parameter encode a ratio at their birth; storing the evaluated number severs the dependency, so changing the base leaves stale semantics with no failing test. Expressions preserve the intent; and an explicit both-directions dependency list (follows / must-not-follow) is what makes a base-parameter change reviewable.
Applies when
- changing gait clock, control frequency, or units
- a reward threshold interacts with a phase/window duration
- config audit finds literals that encode ratios
“feet_air_time 阈值 0.25 是写死的,不跟 ct 走——0.40 时摆动窗 ~0.20s<0.25(恒拉长压力),0.50 时摆动窗 ~0.25s≈阈值(语义翻转)。按比例保原压力:0.25 → 0.3125(=0.625×ct;建议直接写成 0.625 * ct 表达式,一劳永逸)。”
train/WALK_V10_SPEC.md § 3. T —— 慢时钟 (训练侧必做一件) Reward fixes come in causal chains - foot height, then landing impact, then foot spacing
reward-chain-foot-height-landing-spacingPlan reward shaping as a chain, not a point fix: when you patch a degenerate gait behavior, pre-register which adjacent behavior the optimizer will exploit next and watch for it.
Symptom
Three problems appeared strictly in sequence: (1) swing feet lifted too low; (2) after fixing that, feet slammed down - "实际比视频里暴力得多" (far more violent in person than on video); (3) after fixing that, feet drifted too close together and collided.
Context
Each reward fix removed one degenerate optimum and exposed the next. The fix for foot spacing (COM lateral randomization +/-5 cm to force leg spread) itself caused base side-to-side sway, requiring a further foot-to-centerline distance penalty. Lucen had just solved its own foot-height problem (19mm -> 40mm swing height) and logged landing impact and foot spacing as the predicted next two problems.
Change
Chain of additions - (1) penalty when swing foot below 5 cm; (2) landing vertical-velocity penalty at touchdown; (3) COM lateral randomization +/-5 cm, then foot-centerline distance penalty to cancel the induced sway.
Outcome
Reference robot progressed through each stage; each individual fix worked and predictably surfaced the successor problem. For Lucen the chain served as a pre-registered roadmap of what breaks next.
Mechanism
Locomotion rewards are coupled through contact dynamics: raising swing height adds potential energy that must go somewhere at touchdown (impact); penalizing impact and forcing robustness to COM shifts changes lateral support strategy (spacing/sway). The optimizer always exploits the cheapest unpenalized channel, so fixing one channel routes the exploit to its neighbor.
Applies when
- adding a foot-height / clearance reward
- feet slam or landing impact grows after a clearance fix
- feet converge toward the centerline or self-collide
- any single-reward fix to a coupled gait behavior
“抬脚太低 → 加惩罚：摆动足低于 5 cm 就扣分 / 加完之后砸脚 → 抬起来了但落地极猛，"实际比视频里暴力得多" → 加落地速度惩罚 … / 两脚太近甚至互撞 → 先试质心横向随机化 ±5 cm … 有效但引发新问题——基座开始左右摇摆 → 再加足-中心线距离惩罚 … 这三条是串联的：每个修复都会暴露下一个问题。”
Experience.md § 三个问题的解法链 (lines 72-77) Audit rewards by realized contribution (weight x achieved value) - a weight of 2.0 was really paying 0.04
realized-contribution-auditEvaluate a reward table by each term's realized per-step contribution under the current policy, never by its weight column; if a term's realized value is ~0, escalating its weight is a no-op - change the term's structure instead.
Symptom
Foot dragging persisted through repeated weight escalation: feet_air_time had been raised 0.25 -> 1.0 -> 2.0 across versions with no behavioral change, and the training-side comment even recorded the fact ("几乎没有 单支撑相, 是拖着脚蹭") without the fix changing form.
Context
Computing realized per-term contributions in the trained state exposed the economy: feet_air_time contributed weight 2.0 x achieved 0.019 = 0.038 per step, against tracking's +1.20 - lifting the leg earned 3% of what tracking earned, so dragging was the rational optimum no matter the weight escalation. The same table acquitted the energy penalties (total -0.30 negative vs +1.74 positive) that a naive read of weights (-5.0 orientation!) would have blamed.
Change
Fix redirected from "raise the weight again" to "add a term whose realized contribution changes the optimum": a clearance penalty sized so its realized magnitude (~0.018/foot when dragging) is comparable to feet_air_time's, enough to flip the optimum without drowning tracking.
Outcome
With the term economy corrected (plus posture/range fixes), swing height reached 34 mm and tracking 87% by v6; weight escalation of the old term was abandoned.
Mechanism
A reward weight is only a multiplier on whatever the policy currently achieves on that term; when the achieved value is near zero (behavior absent), escalating the weight multiplies near-zero. Optimizer behavior is governed by realized per-step magnitudes, so audits must be conducted in that currency.
Applies when
- a behavior persists despite repeated weight increases
- auditing whether penalties are "too strong" or rewards "too weak"
- sizing a new reward term against existing ones
“把 feet_air_time 权重从 0.25 → 1.0 → 2.0 一路加,但没有加高度项。量级算下来:feet_air_time 权重 2.0 × 实得 0.019 = 0.038,而跟踪奖励是 1.2。抬腿的边际收益只有跟踪的 3%,拖地当然是最优解。… 正项 +1.74,负项 −0.30。能量惩罚不是瓶颈,抬腿没收益才是。”
train/WALK_DIAGNOSIS.md § 决定性证据(#1) / 各项奖励的实际量级 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 A quadratic clearance reward stalled for 3500 iters near target - switch to an indicator on accumulated height
indicator-reward-avoids-gradient-decayWhen a shaped term plateaus near its target, check the gradient profile: replace vanishing-gradient forms with threshold/indicator forms for the final approach, and prefer delta-accumulation over absolute positions to immunize against frame offsets.
Symptom
The quadratic-error clearance term froze at -0.002 from iteration 2000 to 5500 - thousands of iterations with no progress on foot lift.
Context
Diagnosis: a quadratic penalty's gradient vanishes as the error approaches target, so exactly where the last millimeters must be earned the incentive fades to nothing. The replacement (Humanoid-Gym form): accumulate the swing-phase height climb per foot, reward a BINARY indicator |accumulated - target| < 0.01 masked to the planned swing window, reset on contact, weight +1.6 as a positive reward. Two properties: the indicator's incentive is constant until the threshold is crossed (no decay zone), and accumulating height DELTAS makes any constant sole-frame offset cancel automatically - which structurally sidesteps the earlier 0.0585 m zero-point bug ("顺带绕开我先前那个'忘了减 0.0585 导致惩罚恒为 0'的坑").
Change
Clearance reformulated from quadratic penalty on instantaneous height to indicator on per-swing accumulated climb (target 0.03 m by leg-length scaling, weight +1.6).
Outcome
Part of the v5 package under which lift finally moved (v5 29 mm, v6 34 mm vs the stalled 18-24 mm era); the offset-cancellation property removed one whole bug class from the term.
Mechanism
Policy-gradient learning follows the reward's local slope; quadratic shaping concentrates slope far from target and starves it near target, so convergence stalls precisely at the finish line. An indicator pays a constant bounty until the goal is met; formulating on deltas rather than absolutes removes sensitivity to reference- frame constants.
Applies when
- a reward term's value freezes short of target for thousands of iters
- designing clearance/height/precision terms
- reward code depends on absolute link positions
“现行二次型在接近 target 时梯度趋零 —— 这正是 clearance 从 iter 2000 到 5500 卡在 −0.002 不动的原因。… 二值指示在跨过阈值前梯度恒定,没有衰减区 … 累积 delta 让 SOLE_OFFSET 自动抵消”
train/WALK_V5_SPEC.md § 3. clearance 改峰值型(去掉二次型的梯度衰减) Ungated phase shaping made standing 42x more expensive than stepping - and the stepping was cooking the hip motors
moving-gate-42x-stand-taxGate every phase/clock-driven shaping term on the command that justifies motion, price the cmd=0 case explicitly during design - and when a reward flaw is sim-only, book it with trigger conditions instead of operating immediately on a working lineage.
Symptom
At cmd = 0 the sim policy never stood still - it stepped in place and crept 0.98 m per 20 s; on the real robot the same lineage's stepping made hip_roll motors run 20 degC hotter than every other joint (43-48 degC vs 25-28).
Context
The arithmetic closed it: the three phase-shaping terms (joint_pos_ref 1.6, feet_contact_number 1.2, feet_clearance_swing 1.6) are driven by the gait clock with NO command gating, so standing at cmd=0 forfeits 3.32/step of shaping while honest standing earns only 0.078 of tracking - stepping wins 42x. The heat chain: perpetual stepping = perpetual single support = one hip_roll stalled at ~4.3 N*m (25% of torque limit) carrying the torso's frontal-plane moment - two hip_rolls = 90% of whole-machine steady-state I2R; measured stand-vs-step comparison: total heat -71% when actually standing. The suspended test acquitted the actuator (0.21 rad sag -> 0.0008 rad in air) and mechanics vetoed the easy fix ("降 kp 救不了热" - equilibrium torque equals the external load regardless of kp). The fix (moving_gate: hard-gate the three shaping terms on |cmd| > eps) was designed - then DEFERRED by the user because the real robot at the time stood fine: "真机不表现该问题, 为真机不存在的病改奖励表不划算", with written trigger conditions (C-ladder stand row persistently failing, or real robot starting to step/drift at cmd=0) and the known hazard tag (this is exactly the reward-change class that triggers the B-arm signature). When the real robot later DID step and cook, the booked trigger fired and moving_gate moved from debt to to-do with its benefit re-priced: "停止空烧 hip_roll,稳态发热降 ~71%".
Change
moving_gate designed with the gate_by_cmd convention; deferral, triggers, and expected heat recovery all pre-registered instead of patching the reward for a then-sim-only symptom.
Outcome
The cmd=0 stepping went from mystery to closed arithmetic; the thermal measurement (hip_roll +20 degC) quantitatively confirmed the 90%-of-heat prediction; the reward change waited for real-world justification instead of spending a risky revision early.
Mechanism
Clock-driven shaping terms define a perpetual-motion bounty unless gated by command; the resulting idle gait is not a training bug but the table's optimum. Its cost surfaces on hardware as stall-torque heating set by statics (mass x lateral offset), which no gain change can remove - only removing the motion (gating) or widening the stance can.
Applies when
- the policy steps in place or creeps at zero command
- specific joints run hot in idle behaviors
- deciding when a known reward flaw justifies a risky mid-lineage fix
“塑形合计 −3.32/步 … 净: 站定亏 42 倍 … 两颗 hip_roll 4.29 / 4.09 N·m(各占限扭 25%),占全机稳态 I²R 的 90% … 吊挂实测 hip_roll 跟踪误差 0.21 rad → 0.0008 rad … 降 kp 救不了热 … 真机不表现该问题, 为真机不存在的病改奖励表不划算”
train/README.md § C1 FAIL 节 (cmd=0 的 sim/真机分歧记账) / C2 真机 A/B 三b 发热定性 Knee swing collapsed because it directly trades against the slip penalty - price the conflict explicitly and clamp what reward cannot hold
knee-swing-vs-slip-pricingWhen a behavior collapses as another metric improves, look for the term pair trading them and set their price ratio deliberately (with escalation and reverse tripwires pre-registered); where the policy actively spends action budget to undo your target, stop paying more reward and clamp the target space structurally.
Symptom
Knee peak-to-peak swing collapsed across generations - v5 33 deg, v10 26-30, v10b 7-8, v11 6.5-8.6 - and rolling the clock back did not recover it, acquitting the clock; the collapse tracked the gated slip penalty instead: "屈膝与不打滑在当前奖励里直接对抗" - v10b's excellent 93 deg slip was purchased with knee amplitude.
Context
Reward-side flexion fixes had failed three times: raising reference amplitude backfired twice (v9/v11), and v11's deep-squat default was actively fought by the policy - it spent 0.68 of action budget pulling the squat straight ("被策略花 0.68 动作拉直反杀"). v12's design accepted the conflict as real and attacked on two tracks: (1) ECONOMICS - a direct knee_swing_amplitude reward (+0.3, target 0.55 rad, capped at 0.6/step = 55% of tracking), explicitly opposed to the slip penalty by design ("显式对立——这正是设计:v12 就是这场对抗的定价实验"), with an escalation ladder (K +0.3 -> +0.5, then slip -0.5 -> -0.3, one layer at a time) and a reverse tripwire (slip telemetry back at v10 levels -> slip weight to -0.8, accept ~20 deg knee compromise); (2) STRUCTURE - knee target bounds [0.2, 0.9] rad so full straightening is physically impossible (straightest 11.5 deg) and the 0.68 fighting budget is released. A bonus falsifiable prediction was attached: phase-lock strength tracks amplitude (v9_probe 48 deg locked 2.5 Hz; v11 low-amplitude 1.36 Hz unlocked), so if K works, hardware phase-lock should return - one change, two verdicts.
Change
knee_swing_amplitude reward + knee target clamp + pre-registered escalation/reverse levers; the failed reward-side-only approach retired.
Outcome
The lineage was frozen before v12 trained (strategic reset), but the diagnosis stands as the walk line's clearest example of two reward terms trading a behavior between them, with the pricing experiment and structural clamp fully designed and calibrated.
Mechanism
When two terms price opposite aspects of one motion (swing amplitude creates yaw momentum that becomes slip), the optimizer settles wherever the price ratio puts it - patching one side moves the equilibrium, not the conflict. Explicit pricing makes the trade a designed quantity; structural clamps remove the regions where the policy spends budget fighting the designer.
Conflicts
The pricing experiment (K vs slip) was designed and calibrated but never trained - the 2026-08-05 reset suspended v12; the collapse attribution table and the 0.68-action counterattack are measured, the remedy's效果 is untested.
Applies when
- one gait quality degrades in lockstep with another's improvement
- the policy visibly fights a default pose or reference
- repeated reward-side fixes for the same behavior have failed
“膝摆塌在 v10→v10b,头号嫌疑是门控滑移罚(四代实测膝 p2p:v5 33° / v10 26~30° / v10b 7~8° / v11 6.5~8.6°;退时钟没救回 → 非时钟)——"屈膝"与"不打滑"在当前奖励里直接对抗 … 奖励侧修屈膝已三败 … v11 深蹲 default 被策略花 0.68 动作拉直反杀”
train/WALK_V12_SPEC.md § 0. 定位 / 2. K —— 膝摆经济(与滑移罚的对偶) 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) Sort external training advice into adopt / already-have / modify / would-trap by recomputing it on your own config
external-advice-audit-against-own-arithmeticNever apply external tuning advice directly: recompute each claim on your own reward table and probe data, classify it adopt / have / modify / trap, and record why - and verify external citations actually exist.
Symptom
External AI/literature advice for the omni ladder arrived plausible-sounding but was written without knowledge of this robot's actual reward table, contract, and history; following it blindly would have broken single-variable discipline and, in one case, made sidewalk unlearnable.
Context
Before the C ladder, every external suggestion was audited: 3 adopted (ellipsoid command sampling; staged wz bands; command-switch acceptance), 3 already present (unified reward; frame history - the frozen 168-dim 5-frame window; per-100-iter acceptance), 2 modified (stand share kept at 20% to avoid a second variable; back share NOT raised because the probe showed backward works untrained 20/20@67%, so oversampling would only crowd out forward), and 1 flagged as a trap: "start vy very small (0.06-0.15)" - on THIS reward table vy was only an L2 tax, so ignoring a vy=0.06 command costs 0.4% of the vx tracking scale, 28-180x cheaper than ignoring forward, with quadratic shrinkage making small commands weaker still. A separate retrieval-reliability note: two search agents returned fabricated verbatim quotes from arXiv PDFs (2 papers, verified fake and discarded); only HTML/abstract/source-verifiable material was used.
Change
Advice classified only after recomputing each claim with local numbers; the "start small" trap was replaced by adding a gated lateral tracking term (the ladder's only true reward surgery) instead of shrinking the command.
Outcome
The adopted items (ellipsoid modes, staged wz, transition acceptance) entered the ladder; the trap was avoided; one external factual error (calling s1g the mainline start - it was falsified 0/20) was caught. Later, one initially-dismissed item (sigma=0.15 too narrow) turned out right for a different reason than claimed - see cycle-average-tracking-for-gait-quantities.
Mechanism
External advice encodes the advisor's reward table and robot, not yours; the transfer-validity test is whether the claim survives recomputation under your own arithmetic (reward margins, probe baselines, contract freeze). Items that survive become experiments; items that don't become documented traps.
Applies when
- incorporating LLM or literature advice into a training plan
- advice conflicts with locally measured baselines
- an external claim depends on reward-table details the advisor cannot know
“其建议 C4「先很小，vy = ±0.06~0.15」—— 在我们这张奖励表下会让侧走学不起来 … 忽略侧走比忽略前进便宜 28~180 倍，且指令越小激励越弱（平方缩放）—— "先很小"在稳定性上对、在梯度上正好把信号缩没了。”
train/C_LADDER_RUN.md § 1. 外部 AI 训练建议的评估（采纳 / 已有 / 要改 / 会踩坑） 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) Prove a new penalty actually fires - two ways a clearance term silently did nothing
inert-reward-term-auditBefore training with a new reward term, log its realized per-step value under the current policy and confirm it is nonzero where intended - check coordinate zero-points against FK and check who occupies the term's gate; and never weaken the term that creates the states your new term needs.
Symptom
A newly designed swing-height clearance penalty could have trained as a no-op twice over, and the companion advice to lower feet_air_time actively backfired when tried.
Context
Instance 1 (zero-point offset): the proposed code used body_pos_w of the foot link, but that is the ankle_roll_link frame origin, which sits 0.0585 m above the ground even with the foot flat on it - so (0.03 - 0.0585) is always negative and the penalty is永远 0; the 0.0585 offset must be subtracted (verified identical in MuJoCo FK and Isaac). Instance 2 (gate occupancy): the clearance penalty fires only in swing phase; a dragging policy keeps both feet in contact, so the penalty is constantly 0 for exactly the policy it was meant to fix - and worse, any slight lift immediately incurs it, a reverse threshold. Lowering feet_air_time to 0.5 on that advice measurably collapsed air time to 0.0002 (below v3). Corrected understanding: "clearance 是把已有的摆动相抬高, 造出摆动相仍要靠 air_time" - air_time creates the swing phase, clearance raises it.
Change
Fixed the height zero-point; kept feet_air_time as the swing-phase creator with clearance layered on top; both errors documented as corrections to the team's own earlier advice.
Outcome
With both fixed, swing height rose from 22-23 mm (v2) to 29 mm (v5) to 34 mm (v6); the inert-term failure class entered the standing checklist.
Mechanism
A penalty's gradient exists only where its gate is occupied and its argument crosses its threshold; frame offsets shift the threshold out of reach, and phase gates can have zero occupancy under exactly the policy being treated. Terms interact as an ecology - one term must create the states in which another can act.
Applies when
- adding any gated or thresholded penalty (clearance, impact, slip)
- a new term produces no behavioral change at any weight
- body-frame positions are used in reward code
“body_pos_w 是 ankle_roll_link 坐标系原点,平放触地时仍高出地面 0.0585 m。… (0.03 − 0.0585) 恒为负 → 惩罚永远是 0 … clearance 惩罚只在摆动相生效,拖地时两脚始终触地 → 惩罚恒 0;而一旦轻微抬脚就立刻扣分,对正在拖地的策略是反向门槛。… 正确认识:clearance 是"把已有的摆动相抬高",造出摆动相仍要靠 air_time。”
train/WALK_DIAGNOSIS.md § walk_v4 独立验收 — 本文档给的两处代码/建议是错的 Torque caps cannot soften footfalls - impact is falling-mass momentum, only the reward can treat it
landing-impact-not-fixed-by-torque-capsClassify each hardware symptom by the physics that sets it: quantities fixed by ballistic momentum at contact must be treated through the policy's trajectory (reward terms on approach velocity/force), never through actuator caps - and size such penalty weights against your own tracking reward, not a lighter robot's.
Symptom
Footfalls slammed at 1.78x body weight in sim baseline (human walking: 1.2-1.5x); the tempting hardware-side fix was cutting actuator torque limits.
Context
Measured directly: scaling torque limits from x1.0 down to x0.4 left peak landing force essentially unchanged (1.75 -> 1.78x body weight) - the impact force comes from the momentum of the falling mass at touchdown, not from motor effort. The fix has to change the trajectory, i.e. the policy, i.e. the reward: feet_contact_forces penalty above a threshold of 113 N (= 1.2x the 9.58 kg robot's weight), clipped, weight -0.005. The weight was sized locally, not copied: the reference robot's -0.001 would amount to 0.9% of tracking reward on this robot ("策略不会理它" - the policy would ignore it); -0.005 gives 4.4%.
Change
Added threshold-type contact-force penalty (-0.005, threshold 1.2x body weight) as one of v6-minimal's three changes; hardware torque cuts explicitly rejected as a footfall treatment.
Outcome
Landing force 1.72x -> 1.55x by v6 (target <1.5x, missed by 3% - progress booked honestly); the torque-cap dead end was documented so it would not be retried.
Mechanism
At touchdown the ground stops a ballistic mass; the impulse is set by approach velocity and effective inertia, which motors can no longer influence in the final instant. Only earlier trajectory choices (approach velocity, timing) reduce it - and those are selected by the reward, not by actuator limits.
Applies when
- footfall impact or landing noise on hardware
- proposals to derate torque as a softness fix
- importing contact-force penalty weights from another robot
“⚠️ 硬件限扭降不了落脚力 —— 砸地力来自下落质量的动量: 实测 tau ×1.0→×0.4, 落脚力 1.75→1.78× 体重纹丝不动。只有这条奖励能治。… ⚠️ 权重不能用 Pi 的 −0.001 —— 实测在我们身上只占跟踪奖励的 0.9%, 策略不会理它 (Pi 6.94 kg 更轻)。−0.005 给到 4.4%。”
train/WALK_V6_MINIMAL.md § ③ 新增 feet_contact_forces An exponential kernel on instantaneous velocity punishes gait oscillation - track the cycle average
cycle-average-tracking-for-gait-quantitiesReward velocity tracking on gait-cycle averages (or filtered values), not instantaneous samples, whenever the desired behavior oscillates at stride frequency; widening the kernel does not fix a variance penalty.
Symptom
Even while the robot genuinely sidewalked (verified after the metric fix), the Isaac-side tracking reward sat on the ignore-floor: true sidewalk scored 0.178 vs 0.189 for ignoring the command - the reward was mildly punishing the desired behavior.
Context
Sidewalking is inherently oscillatory: per-frame vy std was 0.177 while the tracking kernel width was sigma = 0.15, applied to the instantaneous value. A kernel-width scan showed widening sigma 0.15 -> 0.50 still loses (-0.14 -> -0.07): "指数核惩罚的是方差，而侧步天生带方差" (the exponential kernel penalizes variance, and side-stepping inherently carries variance). Modeling with measured parameters: replacing instantaneous vy with the mean over one gait cycle (0.5 s) flips the margin decisively (true sidewalk 1.888 vs ignore 1.281, +0.607), half-cycle is neutral (+0.006), two cycles adds nothing more. Explicitly flagged as extrapolation pending Isaac-side implementation. This also vindicated a previously dismissed external note (sigma too small) - right conclusion, different mechanism than claimed (variance, not gradient).
Change
Proposed fix recorded: change the tracked quantity from instantaneous vy to a one-gait-cycle running average; widening sigma alone rejected by the scan.
Outcome
Diagnosis complete and quantified; the C4 product shipped via feed-forward before the reward change was implemented, so the cycle-average fix remained a verified-by-model, not-yet-trained change.
Mechanism
E[exp(-(v-c)^2/sigma^2)] decreases with Var(v) even when E[v] = c exactly; a gait's phase-locked oscillation guarantees variance at the stride frequency, so instantaneous tracking rewards structurally prefer standing still at the command mean. Averaging over exactly one cycle removes stride-frequency variance while preserving command-following error.
Conflicts
The cycle-average fix itself is model-extrapolated ("⚠️ 这一条是外推，须在 Isaac 侧实装并复量后才能当结论") - the diagnosis is measured, the remedy untested in training at the time of writing.
Applies when
- tracking rewards for lateral/turn/any oscillation-carrying velocity
- a verified behavior scores below the ignore-floor
- choosing sigma for exp-kernel tracking terms
“侧走时 vy 的逐帧摆幅 std = 0.177，而 track_lin_vel_y_exp 核宽 σ = 0.15，且作用在瞬时值上 … 真侧走（均值 66%，振荡 ±0.18）0.178 | 完全无视指令 0.189 … 真侧走的得分比无视指令还低。… σ 从 0.15 放到 0.50，侧走仍然吃亏 … 把跟踪目标从瞬时 vy 换成一个步态周期（0.5 s）的平均 vy：… 1.888 vs 1.281”
train/C_LADDER_RUN.md § 3n. 二/三 Isaac 训练奖励为何一直坐在「无视底分」/ 修法不是放宽 σ Train with self-collisions ON (filtering nested-link ghost pairs) - the reward wall prevents, the physics makes cheating impossible
self-collision-physics-plus-reward-wallNever train a contact-risk behavior with self-collisions disabled; enable them with an audited filter list for nested/overlapping pairs (zero contacts across a pose sweep), record the fps cost, and keep a calibrated distance penalty as the preventive layer on top.
Symptom
walk_v8 logged 107 frames of leg-on-leg contact while still earning 0.751 tracking score - because training-side self-collisions were OFF, leg clipping was literally imperceptible to the policy ("碰腿在训练里 根本感知不到").
Context
Enabling self-collisions naively is its own trap: an Isaac audit had shown PhysX auto-filters adjacent bodies (base-hip clean for free) but nested links generate ghost forces - calf and ankle_roll overlap 65 mm at the zero pose, producing 12x body-weight phantom forces. The v10 recipe: enable self-collisions, explicitly filter only the two nested pairs (l/r calf-ankle_roll), then run a zero-contact audit at three poses (nominal stand, walk crouch, swing-extreme) requiring contact count = 0, adding any residual pair to the filter and re-auditing; a 500-iter sanity run for NaN and an fps-cost record (measured -8.8%). Redundancy with the reward-side foot-distance wall was argued, not assumed: "N2 离得远(奖励侧预防),SC 碰了疼(物理侧兜底)" - the reward keeps distance at range, the physics makes contact hurt - so the v8-style "clip legs and still score" outcome becomes physically impossible.
Change
enabled_self_collisions=True + 2-pair filter + three-pose zero-contact audit (re-verified at 0.00 N after the later mass update) + fps budget recorded.
Outcome
Leg contact entered the training signal; the audit protocol caught the nested-pair ghost-force hazard before it corrupted training; combined with the calibrated distance wall, later versions held contact = 0 on hardware and in sim.
Mechanism
A hazard absent from the training physics cannot be learned about, no matter the reward; but collision meshes that interpenetrate at rest inject large fictitious forces if enabled blindly. Filtered enabling plus a pose-swept zero-contact audit gives true contact physics with no phantom energy - and layering prevention (reward) with consequence (physics) covers both learning and enforcement.
Applies when
- real robot self-contacts while training scored it healthy
- enabling self-collisions on a model with nested collision meshes
- deciding between reward-side and physics-side fixes for clipping
“PhysX 自动过滤相邻体(base↔hip_pitch 免费干净),幽灵力只在 calf↔ankle_roll(零位嵌套 65mm,12 倍体重)。… 与 N2 互补不冗余:N2 离得远(奖励侧预防),SC 碰了疼(物理侧兜底)—— v8 那种 107 帧互碰拿 0.751 跟踪分的事从此物理上不可能。”
train/WALK_V10_SPEC.md § 4. SC —— 训练侧自碰撞(范围已探明,比想象便宜) After seven patch-generations, freeze the lineage as a regression baseline, fix the structural debts, and retrain from zero
freeze-lineage-fix-structure-restartWhen successive rungs keep trading one symptom for another, ask whether the remaining problems are structural (contracts, latency, sampling, reward-table architecture); if so, freeze the lineage as regression baselines, pay the structural debts, and restart minimal - carrying forward laws and instruments, not weights and weights' patches.
Symptom
The v5-v11 walk lineage had accumulated interacting patches (reward terms, gates, clamps, per-joint scales) faster than it converged on the user's goal; v12's spec itself was superseded before training by an external review's verdict that the remaining problems were structural, not parametric.
Context
The 2026-08-05 status banner records the pivot: the walk profile was rolled back wholesale to v10b parameters, the v5-v11 lineage frozen "只作回归对照" (kept only as regression baselines), and four structural debts were named as prerequisites for a from-zero straight-walk baseline: the action-latency FIFO (fixed with its own test), the ONNX manifest contract, discrete command sampling, and a minimal reward table. The v12 spec - fully designed, partially implemented - was suspended: "本规格挂起,不再按此开训".
Change
Strategy switched from "one more patch generation" to freeze-fix-restart: lineage checkpoints retained as comparison anchors, infrastructure hardened first, then a clean retrain with a minimal reward table (this restart produced the s* generation that later became the real-robot SOTA line).
Outcome
A designed-and-ready training generation was deliberately not run - the review's structural findings outranked sunk design cost; the restart line inherited seven generations of laws (calibrations, gate batteries, falsified fixes) without inheriting their entangled reward table.
Mechanism
Patch lineages accumulate coupled terms whose interactions eventually cost more to reason about than a restart costs to train; the knowledge worth keeping is the laws and instruments (measured plant values, calibrated gates, falsified directions), not the entangled weights. A restart on hardened structure converts the lineage's lessons into a clean initial design instead of another delta.
Applies when
- repeated rungs shuffle symptoms without net progress
- an external review flags infrastructure/contract debts
- deciding between another patch generation and a clean retrain
“同日外部评审定调换路线:冻结 v5~v11 血统(只作回归对照),修结构性问题(latency FIFO 已修 tests/test_action_latency.py、ONNX manifest 契约、离散命令采样、最小奖励表)后从零训直行基线。本规格挂起,不再按此开训。”
train/WALK_V12_SPEC.md § ⚠️ 状态 (2026-08-05) 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) 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 A get-up policy righted itself and sat - three terms paid the seated pose 84% of the return, and the only shaping term that could tell sitting from standing was an exp kernel outputting 5e-5
seated-basin-dead-exp-kernelWhen a policy parks in a degenerate posture, tabulate what each reward term pays that posture against the target (watch contact terms that reward touching rather than bearing load) and evaluate every exp kernel at the error actually observed; a kernel narrower than the real error is switched off, and widening it is a one-variable repair that adds nothing new.
Symptom
R0 converged by iteration 700 and gained 1.7% over the next 2,300; success 0.0% in all four fall categories. Three of four success conditions passed (tilt median 1.0 deg, both feet in contact 99.6%, angular rate low); height passed 0.6% (median 0.204 m against 0.326). The robot knelt in a W-sit: hip yaw +/-47 deg, knees folded to 92% of the hard limit, shins flat, pelvis on the ground, torso vertical.
Context
Minimal reward table: upright (1-g_z)/2 +2.0, base_height linear progress +1.5, stand_pose exp(-||q-q_stand||^2/std^2) x upright gate +1.0 with std 1.0, still +0.5 and feet_on_ground +0.5 both x the upright gate, plus regularizers. The upright gate is a hinge that opens below 30 deg of tilt. The pre-registered fallbacks were then checked against the measured state: tightening the tilt gate was falsified (tilt was already 1.0 deg); a success bonus contradicted the spec's own no-cliff-bounty rule; narrowing the categories was useless (all four converged to the same pose); raising init noise was too weak for a basin this deep. Only a half-rise intermediate state addressed it, and a cheaper repair existed.
Change
R0.1 (user decision, single variable): stand_pose std 1.0 -> 3.0. Not a new term and not a bounty - repairing a declared term that was numerically dead. The two runs' logged env.yaml differ in log_dir and std only.
Outcome
R0.1 58.2% overall (R0 0.0%): supine 91.5%, side 82.2%, mid 55.9%, prone 0/156; knees fully straight; ||q-q_stand||^2 9.99 -> 0.91 and the stand_pose term 4.6e-5 -> 0.90; get-up ~1 s, no re-falls, the curve still rising at the 3,000-iteration cap. Prone stayed at zero and needed a different fix (see prone-dead-end-is-foot-placement).
Mechanism
Sitting earned upright 1.98/2.0, still 0.43/0.5 and feet_on_ground 0.46/0.5 - 3.0 of a 3.57 per-second return - because feet_on_ground asked for contact, not load. The only terms separating sitting from standing were base_height (+0.70/s for standing) and stand_pose, whose kernel at the real 9.99 rad^2 error (75% of it in the two knees) was exp(-9.99) = 4.6e-5 with a gradient near 1e-4. Standing up meant risking 3.0/s to gain 0.70/s while unfolding knees at 92% of their limit under load. With std 3 the same term is exp(-9.99/9) = 0.33 - a live gradient, three quarters of it on the folded knees.
Applies when
- a policy converges early to an upright but low, seated or kneeling pose
- a posture-matching exp term reads ~0 in the training logs
- contact-based rewards saturate while the task metric does not move
“**关键:`feet_on_ground` 只问"触地"不问"承重", 跪坐时双脚确实贴地,照样满分。** 三项 3.0/s = 总回报 3.57/s 的 84%。 … **exp(−9.99) = 4.6e-5** —— 权重 1.0 的项实际输出 5e-5、梯度 ~1e-4, **不是"还没学会",是数值上根本不存在**。 … **R0.1 决定(用户 2026-08-09 定,单变量)**:`stand_pose` 的 `std` **1.0 → 3.0**。 不是加新奖励、不是悬崖悬赏,而是**修复一个已声明但数值失效的项**”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §11 R0 首跑(recovery_r0, 2026-08-09):FAIL —— 翻正了但坐着 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) Four in-lineage attempts to widen the standing stance failed - remove a tax, add a joint-space knife, change the target, add a task-space metric penalty - because the stance was the end state of the get-up path; trained from scratch with the right terms it grew right from day one
stance-decided-by-get-up-pathA posture a skill ends in is shaped by the path the policy takes to reach it; if several single-variable edits to the terminal-phase reward cannot move it, stop editing that phase and retrain with the terminal constraint present from the start.
Symptom
v2_6c stood with its feet 0.159 m apart (task-space) and its hips yawed 45-47 deg the same way, which split on the real robot. Standing-phase reward edits did not move it.
Context
V2.7-A removed the flat-feet tax on compensated stances (stance unchanged); V2.7b added a hip-roll lower-bound hinge (+5 deg in 3,000 iterations, yaw ratchet); V2.8 changed the stand_pose target to a wide flat stance (stance unchanged, yaw not unwound, feet nearly overlapping, mu 0.4 transfer 2%); V2.9 penalized lateral spacing in metres (the policy parked just outside the penalty's gate in a lunge, 0% success). The v2_6c get-up goes through a split and closes the feet together as it rises.
Change
In-lineage stance surgery was formally closed. V3.1 trained from scratch with task-space stance terms present from the first iteration (and, after P1, a positive width band instead of a penalty).
Outcome
V3.1 P1b: lateral stance 0.364 m, foot tilt 0.0 deg, all four categories 100%, MuJoCo mu 1.0 and 0.4 both 100% - with a symmetric toe-out the kinematic audit had not enumerated. P1c (with a yaw guard): 0.355 m, all six acceptance criteria passing, mu 1.0-0.4 all 100%; it became the product.
Mechanism
A converged policy does not rebuild the path that produced its terminal posture; a standing-phase gradient only finds the nearest hack around the posture the get-up delivers.
Applies when
- the final posture of a transition skill is wrong and resists terminal-phase shaping
- repeated continuation rungs produce hacks instead of the intended posture
- deciding between another in-lineage fix and a from-scratch retrain
“窄站距 + yaw 扭是 v2_6c 起身策略(劈叉起身 → 双脚并拢收势)的**结构性 终态**,不是站立段的孤立参数 —— 站立形态由起身路径决定,在血统内只动 站立段奖励改不动它。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §44 结果:V2.8 判 FAIL —— 血统内站姿手术第三次证伪 A binary reward band on the swing knee had zero gradient everywhere below it, so the one-leg policy parked in an unloaded "fake touchdown" that Isaac's 5 N threshold scored as success and MuJoCo showed as real pressing - a capped constant-gradient ramp, retrained from scratch, passed 40/40
binary-band-reward-fake-touchdownShape approach-to-target rewards as capped ramps with gradient from the starting posture, never as bands or indicators; and compare contact-based terms across simulators, because a policy riding just under a force threshold looks perfect in one and wrong in the other.
Symptom
At iteration 1,000 of the first one-leg run the swing foot never lifted: the policy stood with the "raised" foot resting lightly on the ground. In Isaac the contact-match term paid 96% of full marks; the same policy in MuJoCo pressed that foot on the ground for 450 frames.
Context
The swing-leg goal was "shank folded fully back" (knee 1.5-1.95 rad), rewarded as a binary band: +0.8 inside [1.5, 1.95], zero elsewhere. From knee 0.05 to 1.5 rad the term was flat. Contact is judged at a 5 N force threshold, so a foot carrying less than 5 N counts as lifted. The walk line had hit the same disease with a binary indicator (v4) and fixed it with a capped ramp (knee_swing_amplitude).
Change
swing_knee_fold changed from the binary band to a ramp clamp(|q|/1.5, 0, 1) - a constant gradient capped near 86 deg - and the policy was retrained from scratch (V0r1). After the first real-robot try showed the fold still too low, its weight went 0.8 -> 2.0 (V0.1).
Outcome
V0r1 model_2300 passed the full acceptance 40/40 (swing knee 1.72 rad, about 98.5 deg) and was stamped as oneleg_v0.onnx; the cross-simulator disagreement is recorded as the thing that caught the cheat.
Mechanism
A reward that is flat until the target is reached gives no gradient to approach it, so the policy settles for the nearest state other terms reward - here, a foot that satisfies the contact threshold without lifting; a second simulator with different contact force resolution exposes such threshold-riding.
Applies when
- rewarding a posture target with an in-band / out-of-band indicator
- a contact threshold decides whether a foot counts as lifted
- trainer-side contact terms are near full marks while the video looks wrong
“初版二值带 [1.5,1.95] 在膝 0.05→1.5 全程零梯度，策略停在"卸力虚点地"（Isaac 5N 阈下 contact_match 96% 满分 / MuJoCo 同策略 450 帧实压——跨仿真器互证抓作弊）；v4 二值指示同型病，按 knee_swing_amplitude 判例改常数梯度封顶 ramp，从零重训 … **oneleg_v0.onnx = V0r1 model_2300, 40/40 PASS**”
git:Lucen V2@origin/oneleg-line:train/ONELEG_V0_SPEC.md § §4 奖励表 swing_knee_fold 行 / §8 核查单 5 Four consecutive fixes were each continued from the previous fix's degraded state until the user stopped the ladder - "change parameters, don't stack errors" - rolled back to the last good checkpoint and audited the target geometry first
stop-stacking-roll-back-and-auditWhen successive rungs each start from the previous rung's output and the target symptom does not move, stop, roll back to the last good checkpoint and re-derive the next change from an audit; keep the measurements, discard the stacked remedies.
Symptom
After the real-robot splits, the V2.7 ladder tried to widen the stance: a term swap (A), a new stance knife (b), more iterations (甲), a doubled weight (乙). Stance barely moved while hip yaw ratcheted 46.7 -> 49.9 -> 52.5 deg toward its 60 deg limit.
Context
Each rung started from the previous rung's output. The user ruled on 2026-08-11 that things had gone wrong from V2.7-A: go back to v2_6 and rethink which parameters to change instead of stacking errors.
Change
乙 was killed at start and not counted; the product baseline rolled back to v2_6c model_29399; every measurement and law learned on the ladder was kept ("the data is real; what stacked was the treatment"). Before any new training, a zero-training kinematic audit of the stance targets was run.
Outcome
The audit found the stand_pose target itself rewarding the narrow stance (pose-target-geometric-audit) and showed geometrically why the yawed stance could not be widened with flat feet - so 乙 was proven unnecessary without running it. The next in-lineage attempts still failed, which is what established that the stance is set by the get-up path.
Mechanism
A rung continued from a degraded state inherits its compensations, so each new fix answers the previous fix's side effects; the yaw ratchet was the visible trace of that stacking.
Applies when
- three or more corrective rungs in a row without progress on the target metric
- a side-effect metric ratchets in one direction across rungs
- a new rung is being planned from the latest (not the best) checkpoint
“用户裁:"从 V2.7-A 开始就出问题了,应该回到 2.6 再思考如何改变参数而不是 错误叠加。"认账:A 的补丁 → b 的新刀 → 甲的加时 → 乙的加权,每级都从上级 的**退化状态**续(yaw 46.7→52.5° 的棘轮就是叠加痕迹)。 … 数据是真的,叠加的是处置。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §44 方法论裁定(2026-08-11,用户):V2.7 全阶梯叫停,回滚 v2_6 Borrow reward values from other robots as ratios (to tracking weight, to leg length) - never as absolute numbers
transfer-ratios-not-absolutesWhen importing any numeric from another robot's config or paper, identify its natural normalizer (tracking weight, leg length, sqrt(g*L), body mass) and transfer the dimensionless ratio; sanity-check against a same-scale robot when one exists.
Symptom
Published configs offered tempting absolute values (swing height target 0.06-0.08 m, weight -20) that would have been wrong for a robot with half the leg length and a different tracking weight.
Context
The cross-check normalized before transferring: G1's feet_swing_height weight -20 against tracking +1.0 is a 20x ratio, so with local tracking at 1.5 the equivalent is -30, not -20. G1's 0.06 m target on a ~0.70 m leg scales to ~28 mm on the local 0.325 m leg (Humanoid-Gym converts to ~23 mm), confirming the locally chosen 0.03 m and explicitly rejecting copying 0.05-0.08 absolutes. A same-class robot (Menlo's 16 kg) was used to sanity-check feet_air_time (+0.5 vs the local 2.0, flagged over-high). A nondimensional check of the same kind later validated the sidewalk speed target (v/sqrt(gL) = 0.081 vs hardware-verified 0.101 - 0.104 - inside the envelope, conservative).
Change
All borrowed values converted through ratios (weight/tracking-weight, height/leg-length, dimensionless speed) before entering the config.
Outcome
The scaled values worked (0.03 m target matched both the scaling law and measured 22-23 mm baseline); no cross-robot absolute was ever copied raw.
Mechanism
Reward economies are scale-relative (only ratios to the tracking term matter to the optimum) and kinematic quantities are morphology-relative (clearance scales with leg length, speed with sqrt(g*L)); absolutes encode the source robot's scale, ratios encode the design intent.
Applies when
- copying reward weights/targets from open-source configs or papers
- setting clearance heights, speed targets, or impact thresholds
- comparing your weights to published tables
“G1 的 feet_swing_height 是 tracking 的 20 倍(−20 vs +1.0)。我们 tracking 是 1.5,按同比例应为 −30 … G1 目标 0.06 m / 腿长 ~0.70 m,换算到我们 0.325 m 腿长约 28 mm;Humanoid-Gym 换算约 23 mm。故 target 取 0.03 m 是对的 … 不必抄 0.05~0.08 的绝对值。”
train/WALK_DIAGNOSIS.md § 修正 ②(权重放大) / 修正 ③(目标高度按腿长缩放) A time gate that had cured one lineage's rushing made the from-scratch lineage trade away its stance width twice (0.364 -> 0.235 m, 0.355 -> 0.251 m) - its disease was absent there, so the fix was retired and the pre-gate checkpoint shipped
time-gate-vs-wide-stance-retire-the-fixCarry a fix into a new lineage only if its disease is present there; a mechanism that cured one lineage can be net negative in another, and when doubling a term's weight recovers almost nothing, treat the two objectives as structurally in conflict and remove the one whose purpose is gone.
Symptom
V3.1's phase 2 (the 3 s zero gate on standing income, continued from P1b) kept 100% success on every friction level and slowed the get-up, but the lateral stance drifted 0.364 -> 0.235 m and hip yaw crept to 57 deg against its 60 deg limit. With the width band's weight doubled (P2c, after P1c) it drifted again, 0.355 -> 0.251 m, below the pre-registered 0.30 m failure line.
Context
The zero gate had been introduced in V2.5/V2.5b to slow the old lineage's get-up. In V3.1 the rushing was already absent: P1c got up in 0.90-1.06 s with a worst torque ratio of 73.3%, better than the stamped v2_6c, because the full beta curriculum, second-difference smoothing and pull curriculum had cured the violence inside training.
Change
Recorded as a candidate law with two data points - the zero gate and a wide stance are mutually exclusive here - and the zero gate was removed from the V3.1 recipe. P1c (the pre-gate checkpoint) went through the full stamp-level acceptance instead.
Outcome
P1c passed everything: all six criteria, lateral stance 0.355 m, foot tilt P75 2.0 deg, mu {1.0, 0.8, 0.6, 0.4} x 10 seeds all 100%. recovery_v3_1p1c.onnx was stamped and pushed to the robot channel.
Mechanism
The zero gate moves the income toward "stay stable until the end", and under low-friction DR a wide stance has a slip tail, so survival outbids the width band; doubling the band's price bought back only 0.016 m - an auction that does not converge signals structural conflict, not an under-priced term.
Applies when
- porting reward mechanisms from an old lineage into a fresh recipe
- a width, margin or posture metric erodes during a late training phase
- a weight increase produces a negligible change in its target
“**定律候选(二实证):归零门 × 宽站互斥**。 … 加价翻倍只挽回 0.016,竞拍不收敛)。 … **归零门是 v2_5 血统的历史包袱,对 V3.1 配方是净负资产,P2 阶段除名 —— P1c 即终点形态**。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §49 终验(2026-08-14) Removing a hand trim re-exposed the plant offset it had been silently compensating - and a slope scan told bias from sensitivity
hand-trims-hide-plant-offsetsTreat hand-tuned trims as undocumented plant measurements: before deleting one, find what it compensates and re-house that knowledge in the model or the reward budget; diagnose posture errors with a sensitivity sweep to distinguish constant bias from gain problems.
Symptom
After switching from the old hand-trimmed default to the clean geometric zero, the retrained stand policy's only regression was torso lean: 1.8 deg -> 4.1 deg backward.
Context
The old default's ankle-pitch trim (-0.0489/+0.0628) had been pre-compensating a fore-aft COM mismatch; removing the trim removed the hidden compensation, and the posture reward alone was too weak to win it back. A COM sensitivity scan settled what kind of problem this was: sweeping base COM offset -50 to +50 mm gave nearly identical slopes for old and new policies (~0.026 deg/mm) - "不是质心敏感度问题, 是恒定偏置" (not a sensitivity problem, a constant bias). Fix landed in stand_v1b: posture corrected to +0.24 deg while keeping symmetry (<=0.1 deg) and low effort (0.259), disturbance rejection better than both predecessors. Model credibility was checked the honest way: v0's sim prediction at the real COM position (-22 mm) was -2.31 deg lean vs real measured 2.2-3.1 deg - "预测精准命中" - which is what licensed trusting v1b's -0.52 deg prediction. (Side flag from the same file: a sign convention had been documented wrongly in early comments - gravity_base[0] > 0 is forward lean.)
Change
Trims retired in favor of explicit modeling: symmetric geometric default plus a posture-reward budget sized to carry the real COM offset; the offset itself known (real COM ~22 mm behind model).
Outcome
stand_v1b passed acceptance as the standing lineage's final version; the walk-line requirement "加大躯干姿态惩罚权重" was upgraded from suggestion to mandatory, since walking amplifies what standing tolerates (real walk_v1 hit 26 deg lean vs sim 7.4).
Mechanism
Hand trims are plant knowledge stored in the wrong place - invisible, asymmetric, and stale after recalibration; removing them re-exposes the raw plant error. A sensitivity sweep separates the two possible diagnoses (slope change = control problem; parallel offset = constant plant bias), each with a different fix.
Applies when
- cleaning up hand-tuned offsets/trims in defaults or calibration
- a posture bias appears after a default or calibration change
- deciding whether a lean is a COM-sensitivity or constant-offset issue
“两者斜率几乎相同(≈0.026°/mm),v1 只是整体多后仰约 2.4° —— 不是质心敏感度问题,是恒定偏置。成因:旧 default 的踝俯仰 trim(−0.0489/+0.0628)本就预补偿了前后质心偏差,换成零位 default 后这份补偿没了 … v0 在真机质心处(−22 mm)的 sim 预测为 −2.31° 后仰,真机实测 2.2~3.1° 后仰 —— 预测精准命中。”
train/RETRAIN_v2.md § 4b. stand_v1 独立验证结果 / 4c. stand_v1b 验收结果 Freeze the deployment contract, stamp every export, and let an automated checker catch wiring bugs
contract-freeze-and-checkerFreeze and fingerprint the policy I/O contract; ship contract changes as new versioned profiles that leave old artifacts bit-identical; and extend the automated contract checker with every pipeline change, forcing the new path to execute in the check.
Symptom
Contract-level changes (observation layout, action pipeline) are where silent sim/real divergence is born; two real wiring bugs appeared the one time the action pipeline was extended.
Context
The 215-dim observation contract was frozen ("纪元 3，三机 digest" - an era number plus a digest agreed across three machines); proposals that would break it (e.g. a GRU memory) were rejected on contract grounds. Every exported ONNX is stamped and verified with a manifest (onnx_manifest --stamp / --verify), and deployment refuses mismatched combinations. When C4 added the lateral feed-forward, it went in as a NEW profile (omni_ff) leaving the existing omni profile's behavior bit-identical; the checker (check_contract) was extended to force the feed-forward path to actually execute (cmd_vy=0.13) and promptly caught two genuine bugs: (1) re-clamping with soft_joint_pos_limits after the feed-forward (0.23 rad deviation) instead of reusing the parent's clip; (2) indexing processed actions by asset.joint_names instead of the action term's own contract-ordered _joint_names, which landed the feed-forward on the wrong joints (l_hip_yaw / r_ankle_pitch).
Change
Contract discipline as implemented: frozen dims + digest; manifest stamping and refusal; contract changes only via new versioned profiles; checker updated in the same commit as any pipeline change, with inputs chosen so new code paths are exercised.
Outcome
Both wiring bugs caught before any training or deployment ("两个都是 check_contract 当场抓出来的 —— 这次它值回票价"); old deployments provably unaffected by the new profile.
Mechanism
The contract is the only interface the policy and robot share; freezing plus fingerprinting makes divergence detectable, and an executable checker turns "the contract holds" from a belief into a test - but only if its inputs actually drive the new code path.
Applies when
- modifying the action or observation pipeline of a deployed policy
- exporting policies for hardware
- proposals that would change observation dims or history structure
“契约校验抓到的两个真错误（记账，别再犯）：1. 前馈后误用 soft_joint_pos_limits（URDF 限位 ×0.9）重钳 → 0.23 rad 偏差 … 2. 用 asset.joint_names 索引 _processed_actions → 前馈落到 l_hip_yaw/r_ankle_pitch 上 … 两个都是 check_contract 当场抓出来的 —— 这次它值回票价。”
train/C_LADDER_RUN.md § 3j. 契约级改动 / 契约校验抓到的两个真错误 Exponential tracking kernels go flat exactly when the error is largest - pair them with an L2 term for the far field
exp-kernel-needs-l2-far-fieldNever let an exp/Gaussian kernel be the only tracking pressure on a quantity that can drift far from target: pair it with an unbounded (L2) term sized as the "don't diverge" floor, and check which frame the kernel reads.
Symptom
With only an exp-type yaw tracking term (exp(-err/std^2), std 0.25), a robot whose heading had drifted badly received almost no corrective gradient: at error 0.6 rad/s the term evaluates to exp(-0.36/0.0625) = 0.003 - near zero AND flat.
Context
The exp kernel is excellent for fine tracking near zero error but its gradient vanishes at large error - precisely when correction matters most. Fix: add track_ang_vel_z_err_l2 (-0.5), a plain quadratic on the same quantity: "exp 管精细跟踪、L2 管'别发散', 互补". Both terms deliberately read WORLD-frame wz (matching the exp term's source), because this torso sways enough that body-frame wz means are systematically off (measured -0.039 while actually turning +0.152). The same far-field-gradient argument reappears in the v8 risk list: frozen joints could not climb back because their huge error put them on the exp plateau ("远端梯度消失是冻结自锁的帮凶").
Change
Added the L2 companion term at -0.5 alongside the existing exp term (a term that had been in an earlier draft and was lost in a rewrite - itself worth noticing).
Outcome
Corrective pressure restored across the whole error range; the exp+L2 pairing became the house pattern for tracking terms.
Mechanism
d/de[exp(-e^2/s^2)] -> 0 as e grows: the kernel saturates and cannot distinguish bad from terrible. A quadratic's gradient grows with error, covering the far field; summing the two yields monotone corrective pressure with fine shaping near the target.
Applies when
- tracking rewards use exp/Gaussian kernels alone
- a drifted or frozen state fails to recover during training
- designing tracking terms for quantities with large transient errors
“exp 在误差大时梯度趋零, 恰好在最需要纠正的时候失灵。… 误差 0.6 → exp(-0.36/0.0625) = 0.003, 接近零且平坦。… exp 管精细跟踪、L2 管"别发散", 互补。”
train/WALK_V7_SPEC.md § ② track_ang_vel_z_err_l2 −0.5 —— 补 exp 的梯度洞 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） 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)