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
169 cards matching “prone-dead-end-is-foot-placement”.
A nonzero response with the same sign for + and - commands is bias, not ability
same-sign-response-is-yaw-biasBefore crediting any directional skill, test both command signs: response must flip sign with the command; a same-signed pair is a bias to subtract, not an ability to report.
Symptom
Root-selection probe showed nonzero wz "tracking percentages" on turn commands, tempting the read that candidates could partially turn.
Context
During C-ladder root selection, s1e-500's measured yaw rate was +0.084 rad/s for cmd +0.3 and +0.093 rad/s for cmd -0.3 - same sign both ways. The same check on the C2 baseline gave wz+0.20 -> -0.13 and wz-0.20 -> +0.12 (again same sign), while the alternative root s2e_pd-1400 gave +0.16 / -0.16 - opposite signs, i.e. a genuine 16% command response.
Change
Reading corrected and written into the execution sheet: percentages on directional commands are meaningless unless the +cmd and -cmd responses have opposite signs; all three candidates were re-classified as "cannot turn, cannot sidewalk - C2/C3/C4 learn from zero". Acceptance criteria thereafter required "tracking >=50% AND left/right opposite-signed".
Outcome
Prevented crediting turn/sidewalk ability that did not exist; the antisymmetry clause became a standing part of every turn and sidewalk PASS condition (C2, C4, C4-redo levels all carry "且左右反号").
Mechanism
A constant yaw (or lateral) bias projects onto any command's sign convention and shows up as fake fractional tracking; only sign-antisymmetry under command reversal distinguishes a feedback response to the command from an open-loop offset.
Applies when
- evaluating turn/sidewalk/any signed-command tracking percentages
- a candidate shows partial tracking on an axis it was never trained on
- writing PASS criteria for a new directional skill
“C2/C3 那些非零的 wz 百分比不是转向能力 —— 转向+ 与 转向− 的实测同号（s1e：cmd +0.3 → +0.084，cmd −0.3 → +0.093 rad/s），那是恒定偏航偏置。… 三个候选都不会转、都不会侧走。”
train/C_LADDER_RUN.md § 0. 读数纠正（重要，别引错） Single-impulse push recovery is a binary chaotic quantity - cross-machine floating-point divergence can flip the outcome
single-impulse-recovery-is-chaoticNever gate or compare single-event recovery outcomes across machines or domains: evaluate disturbances as survival distributions over phases and seeds, compare longitudinally on one machine, and treat any single-point cliff as unconfirmed until it survives the statistical protocol.
Symptom
Mac evaluation found a hard "0.8 N*s cliff" (0/3 survival) that the training machine flatly contradicted: the identical protocol (0.8 impulse at 8 s, cmd 0.2) survived 3/3 there, and a 0.6/0.8/2/4 cross sweep survived everything.
Context
The verdict became a named lesson ("跨机混沌课文"): whether one specific push at one specific phase is survived depends on a trajectory that diverges across machines from floating-point differences alone - "单次冲量恢复是二值混沌量, 跨机浮点发散可翻结局". The boundary was drawn precisely: the 20-seed statistical gates DO agree across machines (established precedent), but that agreement cannot be extrapolated to single-point recovery tests. Protocol amended: disturbance evaluation uses multiple push phases (8/10/12 s), >=10 seeds, and only same-machine longitudinal comparisons; the Mac-side recommendation built on the unreproducible cliff was not adopted, while its directionally-consistent small-impulse data was kept.
Change
Push evaluation redefined from single-event pass/fail to multi-phase multi-seed statistics, with cross-machine comparison banned for event-level results and allowed for distribution-level ones.
Outcome
A false hardware-relevant "cliff" was prevented from steering the ladder (the s2e push rung decisions were made on same-machine statistics); the chaos lesson was cited again when real push tests were restricted to qualitative cross-domain use.
Mechanism
Perturbation recovery near the viability boundary has sensitive dependence on initial conditions; different BLAS/GPU reduction orders yield different trajectories from identical configs, so a binary outcome at one phase is machine-specific noise. Averaging over phases and seeds restores a quantity whose expectation is machine-stable.
Applies when
- a push/disturbance result differs between machines or sim and real
- designing push-recovery acceptance tests
- a sharp pass/fail cliff appears in a chaotic-regime evaluation
“训练机上 Mac 原协议 (0.8 @8s cmd0.2) 3/3 全活 … 与 Mac 的 +0.8 0/3 直接矛盾。定性: 单次冲量恢复是二值混沌量, 跨机浮点发散可翻结局;统计门 (20-seed 八门) 跨机吻合的先例不能外推到单点恢复测试。协议改判: 抗推评测多相位 (push 时刻 8/10/12s) + ≥10 seed + 只做同机纵向比”
train/README.md § s2e 支线终章 (跨机混沌课文) Slowing the gait clock at deployment is out-of-distribution and backfires - lower the commanded speed instead, or train the knob
cycle-time-override-is-oodAny deployment override must correspond to a dimension the policy was trained to handle; to make a parameter field-adjustable, randomize it in training and observe it - otherwise use the levers inside the trained envelope (commands) and leave the knob alone.
Symptom
Real-robot feedback "walks very fast and unstable" suggested slowing the gait; a deploy-side --cycle-time override existed, making "just slow the clock" a one-flag temptation.
Context
A sim sweep of the override on walk_v6 @cmd 0.3 showed monotone degradation away from the trained 0.40 s cycle: at 0.50 s tilt jumped 7.9 -> 13.2 deg and landing force 1.52x -> 2.24x; at 0.80 s (half speed) clearance collapsed to 3 mm - dragging again - with 20 deg tilt. Meanwhile the legitimate lever, lowering the commanded speed with the clock untouched, improved everything monotonically: cmd 0.1 gave 104% tracking, 6.7 deg tilt, minimum slip - the most stable operating point. The file distinguishes the two "slows" explicitly: lower command = smaller steps at the same 2.5 Hz rhythm; a slower rhythm itself requires retraining - randomize cycle_time (e.g. 0.40-0.65 s) during training and expose it as an observation, and only then does --cycle-time become a field-adjustable knob.
Change
Deployment guidance: never ship a cycle-time override the policy was not trained under; respond to "too fast/unstable" with lower commands; schedule clock variability as a training-time (contract-level) change if a field knob is wanted.
Outcome
The sweep quantified the trap before hardware paid for it (dragging and 2.2x landing force at slowed clocks); cmd 0.1 documented as the stable demo point.
Mechanism
The policy is a function fitted around the training distribution; a deploy-side override moves an input (phase rate) to values never seen, so behavior degrades unpredictably - the knob LOOKS like a capability because it exists in the code, but capability lives in the training distribution, not the interface.
Applies when
- a deploy tool exposes overrides (clock, scale, gains) beyond the training distribution
- hardware feels "too fast/aggressive" and a quick knob exists
- deciding between a deploy-side tweak and a retrain
“0.80s | 1.25Hz | 0.165 | 3mm(拖地) | 20.0° … 慢一半直接崩 … 策略按 0.40 训练, 别的周期属分布外。… 降指令速度才是有效杠杆 … cmd 0.1 是最稳的工作点。… 要节奏本身变慢必须重训 —— 训练期把 cycle_time 随机化(如 0.40~0.65s)并作为观测的一维, 部署时 --cycle-time 就成了现场可调的旋钮。”
train/WALK_DIAGNOSIS.md § 2026-08-01 追加: 调慢步态时钟(--cycle-time)在仿真里是反效果 No parameter tuning on the floor - a failing config retries once, then it is out; anomalies go back to sim
no-field-tuning-protocolHardware time is for executing and measuring the pre-registered matrix, never for tuning: failing configs get one retry then elimination, anomalies get recorded and reproduced in sim, and contract-check bypass flags stay unused.
Symptom
Hardware sessions create pressure to fix problems live - nudge a gain, tweak a scale - which destroys attribution and risks the robot.
Context
The anomaly-handling section of the acceptance sheet is three fixed plays: (1) falls at start -> retry once at the same settings; falls again -> that configuration is eliminated, "不现场调参" (no on-site parameter tuning); (2) limit cycle or motor screech -> stop immediately, record the gain level and the joint, reproduce in sim before any discussion; (3) systematic disagreement with sim -> record it as a finding (hardware outranks sim) rather than adjusting anything to force agreement. Related guardrails elsewhere in the sheet: never pass --allow-unstamped / --allow-plant-drift to bypass manifest checks - if it errors, something real is wrong, stop and look.
Change
Field sessions restricted to executing the pre-written matrix; every fix path routed through sim reproduction and the normal config/rung process.
Outcome
Sessions stayed interpretable (each run matched a documented config) and safety overrides never became habit; anomalies arrived back in sim as reproducible cases instead of half-remembered floor stories.
Mechanism
Field-tuned values are measured under adrenaline on one floor with no logging or baselines - they contaminate the config lineage and are unattributable afterwards; and every bypass flag that skips a contract check converts a designed safety property into an operator promise.
Applies when
- a config fails or oscillates during a hardware session
- someone reaches for a live gain tweak or a bypass flag
- writing the anomaly-handling section of a deployment runbook
“起步即摔 → 换档重试一次, 仍摔则该档出局, 不现场调参。出现极限环/啸叫 → 立刻停, 记录档位与关节, 回 sim 复现再议。… 不要给 --allow-unstamped / --allow-plant-drift —— 三枚 ONNX 都已盖章 … 真要报错说明有别的问题, 停下来看。”
train/REAL_RUN_S2.md § 4. 异常处置 A stronger action_rate penalty cut the median torque demand under the gate and left the p99 at 4x the limit - only a hinge on the pre-clip (computed) torque, weighted by comparison with a peer term, collapsed the tail
tail-torque-needs-hinge-on-computed-demandJudge actuator demand against the deployed limit, read the pre-clip demand (applied torque is censored and gives no gradient on the excess), use an L2 rate penalty for the median and a thresholded hinge on computed demand for the tail, and set a new term's weight from its measured steady magnitude next to a peer term rather than from a back-of-envelope estimate.
Symptom
After R0.5 hip_pitch delivered torque sat at its 12 N*m limit in a typical get-up (demand 119-125% of the limit, p99 4.2x) - zero control margin at exactly the moment modelling error matters.
Context
The 12/17/11 N*m limits are deployment limits written into robot.yaml by set_torque (RS06 at 33% of rated), and simulation uses the same effort_limit - so the gate is judged against them, not the 36 N*m rating (an early reading against the rating was retracted). Applied torque is clipped at the limit - censored data - so demand must be read from computed_torque. With the full-range action contract (hip_pitch scale 1.309, kp 30) a single-step action change of 0.306 already saturates hip_pitch, and action_rate penalizes exactly that change.
Change
R3.0: action_rate_l2 -0.01 -> -0.03 (child-run). R3.1: new torque_headroom = sum relu(|tau_computed|/limit - 0.9)^2, normalized so three motor types share a scale. Its weight was first estimated at -0.1, measured in a 12-iteration run at an effective -0.019 (12x smaller - the estimate had mixed a per-episode-peak p99 with a per-step p99, and at 1% of upright it would have been numerically absent), and set to -0.5 so its steady value (-0.095) matched action_rate's (-0.097).
Outcome
R3.0: sum |da|^2 -64%, success 99.6 -> 100%, delivered median = demand median (the clamp no longer fired in a typical episode), gate PASS at worst 79.9% - but p99 unchanged (hip_pitch 419-432% -> 427-436%). R3.1: p99 hip_pitch -> 148-189% (-57 to -65%), knee 422-439% -> 233-234%, saturation duty -60%, success 100%; the worst joint became hip_roll at 67.7%. Its cost appears in torque-penalty-bought-by-leg-bracing.
Mechanism
A squared-rate penalty presses the whole-episode sum and moves the typical step, not rare spikes; the spikes came from the kp term (large targets while a limb is blocked by the ground - velocity alone could not reach them under vel_limit), and a penalty on applied torque cannot see demand above the clip because every excess sample reads as exactly the limit.
Applies when
- torque demand saturates actuator limits in high-effort skills
- a smoothness penalty improves medians but not peaks
- a new reward term's weight is set by estimate alone
“**必须用 `computed_torque` 而不是 `applied_torque`**:后者被 `effort_limit` 削平, 是删失数据,超限样本全被压成"恰好等于限",对超限部分梯度恒为 0。 … 改按同侪定标取 **−0.5**(稳态 ≈ −0.095,与 `action_rate_l2` 的 −0.097 等量)。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §24 R3.1(torque_headroom 力矩需求越限罚) Audit which joints your imitation term constrains - a task that needs deviation is fighting the reference
imitation-term-scope-auditList which joints your imitation/deviation terms actually constrain and check the new skill's required motion against that list; for balance-coupled joints deliver references as feed-forward residuals, not absolute-position targets - and never assume "reference = 0" is neutral.
Symptom
Sidewalk would not learn despite a dedicated tracking reward; meanwhile the gait-shaping imitation term (joint_pos_ref) computed its error norm over ALL 12 joints while its reference covered only the 6 sagittal joints - roll/yaw reference was constantly 0.
Context
Two prior generations had shown the forward gait itself was taught by joint_pos_ref, not discovered by PPO (v6 halved the shaping and swing height collapsed 35 mm -> 4 mm). So the reference is load-bearing - but sidewalk requires hip_roll to deviate from nominal, and the all-joints norm punished exactly that deviation: "一边悬赏一边罚过程" (posting a bounty while punishing the process). A follow-up experiment (C4-redo3, free_roll=True releasing the 4 roll joints from the norm) raised the regularization headroom 6x -> 27x yet sidewalk stayed flat and released hip_roll wandered, killing other skills - net negative, withdrawn. A --roll-absolute probe showed the converse failure: pinning roll to a clock-driven absolute trajectory drove tilt 6.9 -> 13.7 deg. Conclusion recorded: absolute-position imitation cannot teach actions that must be superimposed on state feedback.
Change
The audit reframed the problem: neither punishing roll deviation nor freeing roll nor absolute roll tracking works; the reference for a balance-coupled joint must be delivered as feed-forward under the policy's residual control (see feedforward-for-phase-locked-skills).
Outcome
free_roll rung: joint_pos_ref term rose 0.887 -> 1.104 (release confirmed effective) but vy stayed flat; regularization hypothesis eliminated by experiment.
Mechanism
An imitation error norm defines a cage: joints inside it are pulled to the reference in absolute position, so any skill requiring systematic deviation is taxed per step; but joints carrying active balance cannot follow absolute references either, since their correct position depends on state. The scope and the delivery mechanism of the reference are therefore design decisions per joint, not defaults.
Applies when
- adding a skill that moves joints your reference sets to zero/nominal
- an imitation or deviation penalty coexists with a new tracking reward
- considering releasing joints from a shaping term mid-lineage
“前进步态也不是 PPO 自己发现的，是 joint_pos_ref 教出来的（v6 砍半塑形 → 抬脚 35 mm 塌到 4 mm…）。而 ref_joint_offset 原本只写 6 个矢状面关节，roll/yaw 参考恒 0 —— 侧走既没被教，roll 一偏离 nominal 反被 joint_pos_ref 扣分。一边悬赏一边罚过程。”
train/C_LADDER_RUN.md § 3e. 为什么首战 FAIL / 3i. 解锁笼子 The shipped checkpoint was chosen by scanning checkpoints on the full gate - neighbours 100 iterations apart failed 1 and 38 cells, late checkpoints degraded - never by taking the last one, and training stopped on signals, not on a schedule
checkpoint-choice-is-a-full-gate-scanChoose a release checkpoint by running the full acceptance battery over a band of checkpoints (including the transfer axis), stop training on measured signals rather than a fixed iteration count, and expect adjacent checkpoints to differ sharply.
Symptom
Gate results moved sharply and non-monotonically between checkpoints of the same run, and the last checkpoint was often not the best.
Context
One-leg V0r1: the 2,000 neighbourhood was best; from 2,500 on the nominal gates degraded (late overtraining); 2,000 itself had one real micro-hop (17.7 mm over 5 frames); 2,300 was all green and shipped. V0r2: failed cells per checkpoint 2,000:19, 2,100:38, 2,200:1, 2,300:3, 2,400:27, 2,500:12, 3,000:18 - 2,200 shipped. The recovery line learned the same from the other side: stopping v2_6 early at a scheduled point left a policy whose re-fall rate had spiked to 9-22% before consolidation healed it ("stop on signals, not on the schedule"), and a continuation's transfer decayed checkpoint by checkpoint while Isaac stayed perfect.
Change
The acceptance rule "scan checkpoints, do not look only at the last one" is written into the one-leg gates (called the S1 discipline); release candidates are chosen from the scan.
Outcome
Both one-leg releases were mid-run checkpoints (2,300 and 2,200) chosen by the full 40-cell battery.
Mechanism
PPO keeps changing the policy after the gates saturate; with no gradient toward the gate's conditions, later checkpoints wander, so gate quality is a noisy function of iteration.
Applies when
- picking which checkpoint of a run to export and stamp
- a run is stopped at a fixed iteration budget
- final-checkpoint results are worse than mid-run smoke tests
“Isaac 侧 S1 纪律: 验收扫 checkpoint，不是只看最后一个。 … 扫描判决: 2000 邻域最优——2500+ 标称面退化（⑤③② 散挂, 晚期过训）, 2000 有一例真微跳（L s100 μ1.2, 17.7mm/5帧）, 2300 全绿。”
git:Lucen V2@origin/oneleg-line:train/ONELEG_V0_SPEC.md § §6 验收门 / §8 核查单 5 与 7 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 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） Choose the fork root by which candidate's shortfalls are recoverable, not by headline score
fork-root-recoverable-shortfallWhen picking a checkpoint to fork from, rank candidates by whether their weaknesses are trainable-back, not by current headline metrics; prefer the candidate whose deficits the upcoming training directly pays for.
Symptom
Multiple candidate checkpoints for the omni-command ladder root, each best at something different: fric-3000 had the best tracking precision (vx 88-91%) and hardened plant robustness; s1e-500 had lower precision (vx 82%) but was the only candidate that could still walk backward.
Context
Root selection ran as a data probe, not a preference vote: 6 candidates x 8 out-of-distribution omni commands x 20 seeds = 960 cells (probe_omni_0808.json). s1e-500 @pw1.0 survived 20/20 in all eight conditions including backward at 67% tracking; the deep-trained fric lineage scored backward 0-3/20 despite better forward precision.
Change
Decision criterion made explicit: list what each candidate exclusively wins at, then ask which of those wins the loser could train back. fric-3000's exclusive wins (precision, plant robustness) are both retrainable - precision is directly optimized by the reward, plant hardening is a planned later pass. s1e-500's exclusive wins (backward plasticity 20/20 vs 2/20, disturbance margin 159/160 vs 125/160, push chirality symmetry 40/40 vs 17/40) had all been shown unrecoverable - push-level rungs failed twice, chirality never recovered even with mirror augmentation on. Root = s1e-500.
Outcome
s1e-500 carried the whole C ladder; its backward skill was preserved through C2/C4 gates (regress budget <=2/20 enforced), and the final C4 product passed a 260-cell battery at 20/20 everywhere.
Mechanism
Training can re-earn anything the objective directly pays for, but capabilities that earlier training destroyed and never restored (plasticity, symmetry, robustness margins) are empirically one-way doors. The information-bearing comparison is therefore recoverability of each candidate's deficit, which the team stated as "独占项的可恢复性正好相反 —— 这就是判据" (the exclusive items' recoverability is exactly opposite - that is the criterion).
Applies when
- selecting a resume/fork root among several checkpoints
- one candidate is more precise but another retains a skill the rest lost
- planning a task-extension ladder from an existing lineage
“fric-3000 赢在精度（vx 88~91%…）与 plant 鲁棒性 → 两样都训得回来…；s1e-500 赢在可塑性（C1 20/20 vs 2/20）、抗扰余量（159/160 vs 125/160）、手性对称（推 ±6 N·s 40/40 vs 17/40）→ 三样都训不回来 … 独占项的可恢复性正好相反 —— 这就是判据。”
train/C_LADDER_RUN.md § 0. 为什么根是 s1e-500（数据，不是偏好） A hardware run without its log is an anecdote - the first real get-up's policy, gain profile and log were never recorded, two CSVs stayed "to be reported", and runbook commands wrote different policies' logs under one copied filename
hardware-log-is-the-attribution-inputMake the log part of the run: name it from the policy and conditions automatically (never by hand-copied filenames), record the policy digest and gain profile inside it, include what the open questions need (torque, joint positions and targets), and treat a session without a collected log as incomplete.
Symptom
The recovery line's oldest open question - whether Isaac or MuJoCo reads torque demand correctly - was waiting on real-robot logs that never arrived, and the verdicts that did arrive could not be tied to files.
Context
deploy_policy writes a CSV per run (--log); the runbook's own analysis snippet reads its joint-position, target and action columns (q_, tgt_, act_), and the recovery hanging checklist asks for torque and joint logs for the whole run, to be compared with simulation. The first real get-up (08-09): policy, gain profile and log "to be recorded later". The first real A/B (08-11): v2_5b's result and both policies' CSVs "to be reported". In the runbook's walking commands, three runs of two different c4 policies log to real_s1e_pw08_teleop_0808.csv, and s1e and s2e_fric runs log to real_c2_700_pw08_teleop_0808.csv - filenames copied from other commands.
Change
None recorded; the spec kept listing the open-loop comparison as waiting for real logs.
Outcome
No real-robot log appears in the recovery spec through §50, so the simulator disagreement stayed unresolved and hardware verdicts stayed unattached to data.
Mechanism
Attribution needs the run's identity (policy digest, profile, conditions) and its signals in one artifact; a filename copied from another command mislabels the file, and a log not collected at the session is rarely collected later.
Applies when
- planning a hardware session whose result should settle a sim question
- log filenames are typed or pasted by hand
- hardware feedback arrives as prose without files
“python tools/deploy_policy.py --policy train/policies/omni_c4_ff800_pj.onnx … --log train/real_logs/real_s1e_pw08_teleop_0808.csv … q,t,a=d[:,c("q_")],d[:,c("tgt_")],d[:,c("act_")]”
RL系统/FOLLOW THIS copy 2.md § Walk 遥控 / S2 / csv 分析片段 (operator runbook, undated) An auto-curriculum ratchet capped out at iter 248 and never engaged - stage difficulty manually or verify engagement
auto-curriculum-engagement-checkPrefer manually staged difficulty with gated transitions; if you use an automatic curriculum, instrument its internal state and alarm when it stops engaging - a saturated curriculum is constant DR wearing a curriculum's name.
Symptom
A curriculum mechanism intended to grow difficulty adaptively (s1f's ratchet) hit its cap at iteration 248 and never bit again - for 96% of the run its effect was equivalent to constant DR, i.e. the curriculum existed in name only.
Context
When external advice suggested graded wz bands (start ±0.15, then ±0.30), the team agreed with grading but explicitly rejected automatic curriculum, citing the s1f episode. The same logic had already been paid for with push grading: ±0.6 failed twice, ±0.3 was feasible - grading matters, but the grade transitions were made by hand at verified checkpoints.
Change
Ladder policy: difficulty staged manually, one band per rung, each transition gated by the acceptance battery; automatic ratchets not used unless their engagement is monitored and demonstrated.
Outcome
Every C-ladder band change (wz ±0.12-0.25 first, wider later) was an explicit, attributable rung; no silent constant-DR-in-disguise runs recurred.
Mechanism
Adaptive curricula couple their own state machine to noisy training metrics; a ratchet that saturates early stops adapting but keeps its name, so the operator believes difficulty is progressing when it is frozen. Manual staging costs more decisions but each decision is observable and reversible.
Applies when
- choosing between auto-curriculum and staged bands for a new skill
- a curriculum's difficulty parameter plateaus early in training
- post-hoc attribution of what difficulty a lineage actually saw
“C2 的 wz 分级（±0.15 → ±0.30，不要一上来 ±0.6）—— 与我们付过学费的 push 分级同形（±0.6 两轮 FAIL，±0.3 才可行）。但必须手动分级，不做自动课程 —— s1f 的课程化棘轮 iter 248 封顶未咬合，96% 时长等价常量 DR”
train/C_LADDER_RUN.md § 1. 采纳 3 条 (C2 的 wz 分级) Set torque limits per joint from measured gait peaks - a uniform percentage is the wrong shape, and training must use the deployed numbers
torque-limit-shape-by-measured-peaksMeasure per-joint torque peaks in the actual gait and set each limit as measured-peak x margin capped at rating; then propagate the same numbers into training and add an automated deploy-time consistency check - never derate by a uniform percentage, never let training assume torque deployment will not grant.
Symptom
A uniform 50% torque derating (18/8.5/7) had piled safety margin on the joints that never use it while cutting the busiest joint below half its measured demand.
Context
Per-joint gait peaks were measured (walk_v5 at cmd 0.3/0.6): RS06 (hip_pitch/knee) uses 5.5-5.9 N*m = 15-16% of its 36 N*m rating - cutting it to 12 is a free safety win; RS02's ankle_pitch runs at 16.2 N*m = 95% of its 17 N*m rating - "它是速度的硬件瓶颈", no room to cut; RS00 measured 36-44%, capped at 11. The resulting shape 12/17/11 replaced the uniform percentage. Sweeps across several limit sets (rated / 50% / 14-17-11 / 12-17-11) produced identical speed, lift, and landing force - within this range the limits do not shape the gait; what matters is consistency: "关键是训练和硬件必须是同一个数", because the exporter fills effort_limit from tau_limit, and a policy trained at rated 36/17/14 "会假设有三倍力矩可用" while deployed at 12/17/11 (exactly the v5 cross-generation inconsistency later suspected in its wild kicking).
Change
robot.yaml tau_limit set to the measured-shape 12/17/11, firmware written to match, and train/isaac_values.py regenerated so training sees the same limits; the deploy tool self-checks limits against robot.yaml on every run.
Outcome
Free safety margin captured where demand is low, the real bottleneck joint left at rating, and the train/deploy torque worlds unified with an automated consistency check.
Mechanism
Torque demand is grossly unequal across joints in a gait (15% vs 95% of rating here); a uniform percentage misallocates the safety budget by construction. And since the trainer treats effort_limit as a plant truth, any train/deploy mismatch is an invisible plant gap of exactly the mismatch ratio.
Applies when
- choosing safety torque limits for a legged platform
- training-vs-deployment actuator limit audit
- one joint runs near rating while others idle
“曾用统一 50%(18/8.5/7)是错的形状: 把余量堆在用不到的 RS06 上, 却把 ankle_pitch 砍到需求的 52%。… RS02 在 0.6 m/s 已用到额定 95%, 它是速度的硬件瓶颈 … 实测多组限幅 … 完全一致 —— 限幅在这个范围对步态零影响, 关键是训练和硬件必须是同一个数。… 若训练仍按额定 36/17/14, 学出的策略会假设有三倍力矩可用。”
train/WALK_V6_MINIMAL.md § 3. 训练侧必须同步的一件事 mj_objectVelocity returns inertial-principal-axis frame - one API assumption poisoned eval and observations for a whole line
body-frame-velocity-api-auditVerify every frame-sensitive API against a hand-computed truth (rotate raw qvel yourself, or command a known world velocity and check where it lands) before trusting any evaluation or observation built on it - especially when a model's inertial frame is rotated from its body frame.
Symptom
Sidewalk vy read ~0 under every condition; separately, whole-policy performance was mysteriously mediocre in sim2sim while training-side numbers looked fine. Four training rungs were declared FAIL partly on these readings.
Context
base_link's URDF inertial frame is rotated 90 deg about x relative to the body frame (iquat = [0.7071, 0.7071, 0, 0]). mj_objectVelocity(flg_local=1) rotates into ximat - the inertial principal-axis frame - not the body frame, and returns center-of-mass point velocity, not body-origin velocity. Consequences measured: the "vy" column was actually vertical velocity vz (walking at cmd 0.25: old reading +0.0093 vs true -0.0424); the angular velocity fed to the policy in sim2sim was [wx, wz, -wy] - a different quantity than Isaac and the real IMU provide. RMS check over 8 s of walking: y/z axes swapped between v6[:3] and the qvel truth.
Change
Fixed sim2sim and both probes to compute ang_b = qvel[3:6] and lin_b = xmat.T @ qvel[0:3] (identical quantity to Isaac's root_ang_vel_b / root_lin_vel_b), with a standalone reproduction script (frame_bug_repro_0809.py).
Outcome
Re-scoring the "failed" C4 lineage under correct coordinates reversed the verdicts: c4r4 checkpoints showed vy 80-126% tracking (old reading: +/-2%) and vx+0.30 at 91-95% where the old metric said 0/5 - the bad frame both mis-measured vy and, via corrupted policy observations, systematically depressed all measured performance. Final product passed 260/260 cells.
Mechanism
A simulator API's frame convention is part of the observation contract; when the model's inertial frame is rotated relative to the body frame, frame-agnostic use of a "local" velocity silently permutes axes. Feeding a policy an axis-permuted angular velocity is an observation corruption that degrades behavior everywhere, not just on the axis being studied.
Applies when
- building or auditing a cross-simulator evaluation harness
- one measured axis reads near-zero under all conditions
- sim2sim scores are inexplicably worse than training-side metrics
- URDF/MJCF inertial frames are rotated relative to body frames
“base_link 的 iquat = [0.7071, 0.7071, 0, 0] … mj_objectVelocity 用的是这个 … 喂给策略的 base_ang_vel 是 [wx, wz, −wy] —— MuJoCo 侧观测与 Isaac / 真机 IMU 不是同一个量；vy_mean 报的是竖直速度 vz —— 前进 cmd 0.25 时旧读数 +0.0093，真值 −0.0424。”
train/C_LADDER_RUN.md § 3l. ⚠️ mj_objectVelocity 读的是惯性主轴系 / 3m. 一 bug 坐实 The action-delay was implemented as lerp - beyond one step it extrapolated BACKWARD, so a whole lineage trained on a fictitious actuator
latency-lerp-reverse-extrapolationUnit-test plant-model code (delays, filters, randomizers) against hand-computed truth across its FULL configured range, not just the nominal case - a delay must be a queue, and any interpolation used outside [0,1] is a silent plant corruption that training will faithfully absorb.
Symptom
Every walk/stand model up to v11 had been trained on a silently wrong plant: the action-latency implementation lerp(cur, prev, lag) is only an interpolation for lag <= 1 - at lag 3 it computes 3*prev - 2*cur, a REVERSE extrapolation. With the configured (0, 0.06) s at 50 Hz (lag in [0,3]), about 2/3 of environments were adapting to actuator dynamics that do not exist.
Context
Listed as evidence item #1 for the full restart: "全部旧模型训在错误 plant 上" and the head suspect for the real robot's wild kicking. The fix replaced it with a true FIFO delay line (commit 7f13793) plus its own regression test (tests/test_action_latency.py) - but every exported ONNX predated the fix, which is part of why the lineage was frozen rather than patched.
Change
Delay implementation rewritten as an honest FIFO with unit tests; the restart baseline trained on the corrected plant from day one.
Outcome
A generation-scale training investment was revealed to have a corrupt plant underneath; the class of bug (plausible-looking math that silently changes meaning outside its valid range) got a permanent test.
Mechanism
lerp(a, b, w) leaves the segment for w > 1; used as a delay it fabricates high-gain inverted dynamics precisely in the largest-delay draws, so the policy learns compensation for an actuator that cannot exist - and DR then trains robustness to the artifact rather than to reality. No training metric can catch this: the sim is self-consistent, just wrong.
Applies when
- implementing or auditing action delay / filtering in a trainer
- a lineage behaves as if compensating dynamics nobody modeled
- deciding whether old checkpoints are salvageable after a plant bug
“动作延迟旧实现 lerp(cur, prev, lag) 在 lag>1 时是反向外插(w=3 → 3·prev−2·cur),配置 [0,0.06]s@50Hz 即 lag∈[0,3],约 2/3 env 在适应不存在的执行器动态。7f13793 已换真 FIFO,但所有 ONNX 均训于修复之前 —— 真机"乱踢"的头号嫌疑。”
train/OMNI_V0_SPEC.md § 0. 为什么从零 (1) Add a single-point-suspension test to acceptance - the ground is a free stabilizer that hides divergence
suspension-probe-removes-free-stabilizerInclude at least one acceptance condition that strips the environment's free stabilization (suspension, or equivalent) - the sim-passing policy that fails on hardware is often failing a condition the battery never posed.
Symptom
walk_v5 looked healthy in every on-ground sim test yet diverged on the real robot - the acceptance battery had never measured a condition that would have revealed it.
Context
The battery gained a single-point-suspension probe (robot hung, feet free): measure torso tilt while the policy runs without ground contact. v5 scored 45.9 deg mean tilt suspended - wildly unstable - which the file calls "最灵敏的失稳探针(拿掉地面这个免费稳定器)": ground reaction forces passively stabilize a marginal policy, so on-ground metrics saturate long before the policy's internal balance is actually sound. v6 halved it (23.0 deg, target <10 deg) - progress visible on a scale where on-ground numbers showed nothing.
Change
Suspended-tilt added as a standing acceptance row; run under the honest contact parameters battery (accept_v2 with measured condim 4 / torsional friction 0.035), under which v5 correctly FAILS in agreement with the real robot.
Outcome
The sim battery's verdict on v5 flipped from pass to fail, matching hardware; suspended tilt became the discriminating metric between v5 and v6 (45.9 vs 23.0 deg) when ground metrics differed little.
Mechanism
Contact with the ground closes a stabilizing feedback loop the policy gets for free; removing it exposes the policy's own attitude control authority. A metric measured only in the assisted condition cannot rank policies by the unassisted quantity that hardware will actually demand during perturbations and flight phases.
Applies when
- sim acceptance passes but hardware diverges
- designing an acceptance battery for a legged robot
- two candidates tie on ground metrics
“单点吊那条是最灵敏的失稳探针(拿掉地面这个"免费稳定器"), v5 在地上一切正常却在真机发散, 就是因为验收从没测过这个工况。”
train/WALK_V6_MINIMAL.md § 5. 验收 The recovery line's real-robot verdicts live in three places that disagree - the spec's "fairly stable, first real stand-up" for v2_6, an undated runbook note that only v3_1p1c works, and a first run whose details were never recorded
write-hardware-verdicts-backA hardware verdict is a dated entry in the authoritative ledger - policy file and stamp, gain profile, floor, battery, tries, log file, what was seen - written back the same day; a note in a command file is a pointer, not a verdict, and a newer verdict that contradicts an older one must say so.
Symptom
Asked "which recovery policy works on the real robot", the sources give different answers, and none of them carries the conditions of the test.
Context
08-09: the first real run was stopped as violent and dangerous; which ONNX, which gain profile and whether a log existed were marked "to be recorded" and never were. 08-11: v2_6 was "fairly stable", the line's first real get-up, with splits after standing; v2_5b's result and both CSVs were "to be reported". 08-14: v3_1p1c was stamped and pushed, with "the real first test still needs the user present"; the spec records no hardware result for it. The operator runbook (undated) puts above the v2_6 and v2_5b floor commands the note that none of the recovery policies below work, only recovery_v3_1p1c - a verdict never written back into the spec, with no date, floor, battery, number of tries or log attached.
Change
None recorded in the sources; this card records the gap.
Outcome
The line's authoritative record ends with v3_1p1c as the product awaiting its first real test, while the operator's note implies it is the only one that works and that v2_6 (recorded as a success) does not.
Mechanism
Verdicts given at the robot travel by word of mouth and command-file comments; without a record carrying the conditions, a later reader cannot tell a changed verdict from a changed floor, battery or stack.
Conflicts
§43 (2026-08-11) records v2_6 as the first successful real get-up ("fairly stable"); the undated runbook says every recovery policy except recovery_v3_1p1c does not work; §49 (2026-08-14) says v3_1p1c's first real test was still pending. The runbook's claim has no date and was never written back to the spec, so it cannot be ordered against §43.
Applies when
- choosing which policy to deploy from an operator's notes
- a hardware session ends without a written result
- two documents disagree about what worked on the robot
“下面的recovery都不行 只有recovery_v3_1p1c.onnx”
RL系统/FOLLOW THIS copy 2.md § #### Recovery Policy (operator runbook, undated) Ideal PD is not enough - add a delay buffer and fit armature/friction/delay per joint
actuator-delay-buffer-fittingNever ship ideal PD to hardware: add a measured delay (in control steps) and per-joint armature/friction fitted from step and sine responses, and treat remaining actuator mismatch as your standing largest sim2real residual.
Symptom
Standard ideal PD actuator model transfers poorly; sim assumes targets take effect instantly and joints reach arbitrary acceleration.
Context
A developer with a successful on-hardware Isaac Lab biped modified the actuator model in two ways and calibrated it against the real robot: step-response plus sine-sweep tests (positive step, negative step, sine tracking), overlaying sim curves on measured curves and hand-tuning.
Change
(1) Delay buffer: action targets take effect after a uniform 6 time-step delay on all joints; (2) acceleration limiting so the actuator cannot reach arbitrary acceleration; (3) per-joint fit of armature / friction / delay - different joints genuinely needed different values.
Outcome
Hip joints fit worst, knee best; the developer rated the result "not perfect, the best I could do" and still listed actuator-model improvement as next work - i.e. even the fitted model remained the dominant residual.
Mechanism
Real actuation is a lagged, bandwidth-limited system; a delay buffer and acceleration cap are the two cheapest structures that reproduce its phase and magnitude response. Per-joint differences come from differing load, wiring, and friction states, so a single global constant underfits.
Applies when
- actuator model in sim is ideal PD with no delay
- step-response of real joint visibly lags or overshoots the sim's
- budgeting which sim2real gap to attack first
“标准 ideal PD actuator 不够用，他改了两处：延迟缓冲：目标不是立即生效，全部关节统一 6 个 time step 延迟 / 加速度曲线：执行器不能瞬间达到任意加速度 … 用 armature / friction / delay 三个参数逐关节拟合，标定方法是阶跃响应 + 正弦扫描 … 髋部关节偏差最大，膝关节最好。”
Experience.md § 执行器建模 —— 最值得抄的一条 (lines 50-59) Two machines, one configuration - every gain, offset and torque limit changes only in robot.yaml, whoever edits pushes at once, both checkouts show the same commit before the robot moves, and a pulled policy file is size-checked and synced before power-off
two-machine-config-disciplineTreat the robot's configuration as a versioned artifact with one source of truth, push every change immediately, verify identical commits on every machine before a hardware session, keep hardware limits in the repo and the firmware in sync in both directions, and verify transferred model files (size, digest) before running them.
Symptom
The robot's onboard computer runs the bridge and deploy scripts from its own checkout while training and analysis happen on other machines; a hot fix left on one side, or a half-written file, silently makes the robot run something other than what was evaluated.
Context
The operator runbook's "wall version" of the two-machine discipline: configuration changes only in robot.yaml (calibration offset/sign, gains, torque limits), committed and pushed from the Mac, pulled on the robot, bridge restarted; code is not edited on the robot, and if it is, it is committed and pushed on the spot - nothing unpushed overnight; 30 seconds before every real-robot session both checkouts must show a clean status and the same last commit hash; changing tau_max requires writing the motor's limit_torque too (and the reverse); re-zeroed motors require re-measuring offsets. The recovery line added: after pulling on the robot, check the ONNX is not zero bytes (a lesson from a corruption incident on 08-12) and sync before cutting power; the recovery and main lines are separate worktrees, each pulled with --ff-only.
Change
Operating rules, pinned on the wall and repeated in the hanging checklists ("git pull, both machines on the same commit").
Outcome
The sources record the rules and the incident that produced the size check; they do not record a count of sessions the rules caught.
Mechanism
A policy is evaluated against one configuration; any divergence between the machines, or a truncated file, turns a hardware result into a result about an unknown configuration.
Applies when
- a robot's onboard computer and a workstation both hold the configuration
- someone hot-fixes code or gains on the robot
- model files are copied or pulled to the robot before a session
“改配置只改 robot.yaml(标定 offset/sign、增益、限扭全在里面)→ Mac git commit + push → NX git pull → 重启桥。 … 谁改完谁立刻推,永远不留未推送的改动过夜。 … 每次上真机前 30 秒检查:两边 git status 干净、git log -1 哈希一致。 … 铁律不变:改 tau_max 必须同步写电机 limit_torque(反之亦然);电机重新标零后 offset 必须重测回填 yaml。”
RL系统/FOLLOW THIS copy 2.md § ② 双机维护纪律(贴墙版) Pre-register the ladder's risks and how each future result will be read - before training
preregister-risks-and-fork-readingsBefore a training ladder or risky rung, write the risks, the stop rules, and how every plausible outcome will be interpreted - then do not edit them after seeing results.
Symptom
Without pre-registration, ladder results get rationalized after the fact; the team had already seen post-hoc reads go wrong and adopted written pre-commitment.
Context
The C-ladder execution sheet opens with three numbered pre-registered risks: (1) S2 plant robustness will not carry into omni - a full S2 redo is budgeted from the start; (2) the s1e recipe has a collapse valley at iter 1500+ (recorded twice), so every rung runs a watch_ckpt --every 100 smoke loop with stop-on-degradation; (3) the root's OOD survival edge may partly be "it is slower / commits less" - with the reading fixed in advance: if C1 training raises tracking while survival drops, the edge was bought with slowness, and root selection reopens. Later rungs went further, pre-registering a full result-to-conclusion table for the A/B arms ("预注册读法（事后不改）") and even pre-registering the author's own doubt that a level would fail and what its failure would prove.
Change
Standing practice: before each rung, write down (a) known risks with their mitigations, (b) the interpretation of each possible outcome, (c) stop criteria - all frozen before the run starts ("开训前写死，事后不许改").
Outcome
When arm-A/arm-B and redo results arrived, conclusions were read off the pre-registered table instead of argued; a predicted-likely-FAIL level (C4-redo3) was still run because its pre-registered value was eliminating the regularization hypothesis - which it did.
Mechanism
Pre-commitment converts each training run into a decisive experiment: outcomes falsify or confirm named hypotheses instead of being absorbed into a story; it also makes negative results valuable (a FAIL that eliminates a hypothesis advances the search).
Applies when
- launching a multi-rung training ladder
- running an A/B fork whose outcome will drive a fork/root decision
- a rung is expected to fail but is run for its diagnostic value
“三条预注册风险 … 若 C1 训后跟踪提上去而存活掉下来，说明这条优势是速度买的、不是通用性 → 那时重开选根。”
train/C_LADDER_RUN.md § 0. 三条预注册风险 Push DR helped one lineage and hurt another at the same dose - robustness budget is conserved and gets borrowed, not created
push-dr-conditional-budget-conservationBefore opening a disturbance-DR rung, measure whether the untrained policy already meets the spec; if training it anyway, expect the benefit to be conditional on the lineage's existing DR load, grade the intensity, and audit retained margins - budget spent elsewhere will be borrowed back.
Symptom
The push rung's outcome flipped with the lineage: direct +/-0.6 m/s push failed outright on first attempt (base walking collapsed - kd1.2 scan 0/3 from iter 3300, sim2sim self-falls with pushes OFF - no PASS point existed); staged +/-0.3 then gave the narrow-kd single-working-point lineage real gains (push survival 1/5 -> 4/5) while the SAME dose made the dual-working-point balanced-band lineage WORSE (20-seed survival 18 -> 12/20 plus across-the-board push regression).
Context
The four-ladder verdict ("四梯定案", s2e/s2f at both intensities) named the pattern: "push DR 收益条件性" - the benefit is conditional on how much robustness budget the lineage has already spent. The law candidate: "DR 总预算守恒, 平衡带鲁棒性从抗扰余量借" - total DR budget is conserved; a lineage already covering a wide plant band pays for push tolerance out of its disturbance margin. Both S2 ladders therefore closed at the friction rung, with the decisive numerator: untrained push tolerance already met the 4-6 N*s requirement, so the rung was not needed at all ("⑥ push 不训(收益条件性,免训 ±0.6 已达 标)"). The same accounting later justified the C-before-S2 ordering ("push/μ 两轮已实证 DR 预算有限且会被重分配") and trimmed the second S2 pass to three rungs.
Change
Push removed from the standing ladder; graded intensity retained as the method IF a lineage ever needs push training; "does the untrained policy already meet the disturbance spec" instituted as the first check before opening any disturbance rung.
Outcome
Two rungs (push, ground mu) deleted from the second S2 pass on measured grounds; the ladder's real yield was re-stated honestly as precision, not robustness (speed gate 0 -> 20/20, zero-command drift 0.98 -> 0.06 m, but push 159 -> 125/160).
Mechanism
A fixed-capacity policy allocates representation and margin across the training distribution; adding a disturbance axis to a lineage that already spans a wide plant family forces reallocation - the new tolerance is bought with existing margins. Lineages with narrow plant coverage have free budget, so the identical DR dose lands as gain. Benefit is a property of (dose x lineage state), never of the dose alone.
Applies when
- proposing push/perturbation training on a hardened lineage
- the same DR rung helped one lineage and hurt another
- accounting where a ladder's robustness gains actually came from
“push DR 收益条件性 —— s2e⑥a (单工作点血统 kd 窄带) ±0.3 得抗推 1/5→4/5; s2f⑥ (双工作点平衡带血统) 同档反而 20-seed 存活 18→12/20 且抗推全面倒退。规律候选: DR 总预算守恒, 平衡带鲁棒性从抗扰余量借。两阶梯均以 ⑤ 摩擦级收官 … 抗推 4~6 N·s 免训已达标。”
train/OMNI_V0_SPEC.md § 4. ⑥ push 四梯定案 (2026-08-07) A joint-velocity penalty meant to slow the get-up cut joint speed 16% and left the get-up time unchanged - the knob never moved the variable, so the idea it was meant to test stayed untested
dof-vel-penalty-is-not-a-pacing-knobBefore reading a result as a test of an idea, check that the knob actually moved the independent variable; velocity regularizers smooth a schedule they do not set, and a schedule driven by per-step task income moves only when that income's time structure does.
Symptom
The get-up took 0.6-0.9 s in Isaac with large torque demand; the user proposed getting up more slowly so less torque would be needed.
Context
The idea had support in the accounts: the acceptance bound is an upper bound of 5 s (5-8x margin), the quasi-static squat path peaks at 25% of the limits, and rolling over needs no momentum. R3.2 raised dof_vel from -1e-3 to -5e-3 as the single variable.
Change
dof_vel -1e-3 -> -5e-3 (child-run from R3.1).
Outcome
Get-up medians moved +0.02-0.04 s (noise); raw joint velocity -16%; torque demand median got worse (hip_pitch 46-48% -> 63-67%) as the new term competed with torque_headroom on the same joints; MuJoCo 98 -> 96%. Verdict FAIL on the knob, not on the idea, and the rung was not adopted. When pace was later attacked through the income's time structure (V2.5/V2.5b), the MuJoCo get-up moved into the 3.5-4.5 s design band.
Mechanism
The pace was set by base_height_progress paying for every step spent high (stand earlier, earn more); a velocity regularizer only smooths motion along the same schedule and does not change when the robot stands up.
Applies when
- trying to make a skill slower or gentler with smoothness penalties
- an experiment's primary metric did not move and a verdict is being written
- two penalties act on the same joints
“**关键判读:`dof_vel` 罚只把关节速度压了 16%,而起身用时一点没变。** 也就是说**这一级根本没有把"慢下来"这个自变量推动起来** —— 所以它**不构成对 用户假说的检验** … 起身节奏由 `base_height_progress` 的逐步计酬决定(早站起来就多 拿),速度正则只在同一条时间轨迹上把动作抹匀,不改变何时站起来。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §25 R3.2(dof_vel −1e-3→−5e-3,慢一点起身) A stand gate judged by survival passes a robot that wanders a meter - judge posture instead
stand-gate-posture-not-survivalFor every gate, ask what behavior the metric is a proxy for and validate its ordering against real observations; replace metrics whose ordering disagrees with reality, and demote them explicitly rather than silently.
Symptom
Real robot "standing" drifted 0.5-1.4 m across the floor while the sim stand gate scored a clean 20/20 - because the gate's metric was episode survival, which wandering does not violate.
Context
C2's stand condition was originally written as "survival regression <=2/20". A real-robot counter-example on 2026-08-08 forced the re-judgment: wandering robots survive. Cross-checking candidate sim metrics against real-robot feel showed max-tilt median ordering agreed with hands-on ranking, while displacement ordering was actually OPPOSITE to real impressions - so displacement was demoted to a reference quantity, not a gate.
Change
Stand PASS criterion rewritten from survival to posture: "stand tilt max median <= root baseline +1.5 deg"; displacement kept only as reference. Applied to all subsequent rungs (C4 and redo levels inherit it).
Outcome
Later rungs gated stand on tilt (e.g. C4 product: 7.7 deg vs parent 7.1 deg, +0.6 deg PASS); the wandering failure mode became visible to the battery instead of hidden by survival.
Mechanism
A gate metric is a proxy for an intended behavior; survival is a proxy for "did not fall", not "stood still". Metric choice must be validated against ground truth (real-robot feel/measurement), and a proxy whose ordering disagrees with reality on real data is worse than no metric - it steers selection backwards.
Applies when
- writing PASS conditions for stand/idle/hold behaviors
- a gate passes policies that visibly misbehave on hardware
- choosing between candidate metrics for an acceptance battery
“stand 必须用位姿判，不能用存活判（2026-08-08 真机反证改判）：站着乱走 0.5~1.4 m 时存活照样 20/20 —— C2 的 stand 条件原写「存活退化 ≤2/20」，选错了指标。改为: stand 倾角 max 中位 ≤ 根基线 +1.5°（倾角与真机手感排序一致；位移排序与真机相反，降为参考量）”
train/C_LADDER_RUN.md § 3b. PASS 条件 ⚠️ stand 必须用位姿判 An edge-triggered landing penalty missed the tail and fired after the harm - penalize overspeed continuously inside the contact window
penalize-tail-before-touchdownPenalties aimed at impact/violation events must (a) price the excess over a threshold, not the mean, and (b) be active on the approach (state-gated window), not triggered by the event - check your control rate can even see the event you are penalizing.
Symptom
The v7 landing penalty (vz^2 on the contact-force rising edge, weight -10) did not bite: landing-velocity 95th percentile stayed at 2.61 m/s against a 0.3 target.
Context
Two structural faults were identified: (1) it penalized the MEAN over sparse events - many soft landings dilute the occasional violent slam, while the damage (GRF peaks, motor peak load) lives in the tail; (2) it fired AFTER touchdown - at 50 Hz evaluation the rising edge is aliased by physics decimation, so the read vz is often the already-decelerated post-impact value: underestimated, and with no shaping gradient before contact. Replacement: continuous penalty while the sole is inside a height gate (h < 0.03 m): relu(-vz - 0.30) - only the excess over an allowed approach speed is penalized (tail only), and gradient exists for several frames BEFORE touchdown. The sole-height computation again subtracts the 0.0585 m link offset ("WALK_DIAGNOSIS 坑#1, 别再踩"); the edge-triggered version was kept as a diagnostic only.
Change
feet_landing_vel reformulated: edge-event vz^2 -> in-window relu(-vz - v_ok) with v_ok 0.30 (conservative vs the sqrt(L)-scaled human value ~0.19, to be tightened after passing), h_gate 0.03, weight unchanged -10.
Outcome
The failure analysis of the first form was written before the second was trained; the v_ok escalation path (0.30 -> 0.45 if the robot becomes afraid to land) was pre-registered in the risk table.
Mechanism
Sparse-event mean penalties optimize the average case while the constraint is a quantile; and any penalty evaluated only at/after a discrete event gives the optimizer no gradient along the approach trajectory that determines the event. A state-gated continuous excess penalty fixes both: it prices only violations and shapes the approach.
Applies when
- impact/landing penalties fail to move tail percentiles
- a penalty is triggered by contact edges at a coarse control rate
- designing constraint-style penalties for rare violent events
“罚的是均值路径:上升沿是稀疏事件 … 大量软着陆稀释偶发猛砸;而伤害在尾部 … 罚在触地后:50 Hz 评一次,上升沿被物理 decimation 混叠,读到的 vz 常是撞完已减速的值——既低估,又没有触地前的塑形梯度。”
train/WALK_V8_SPEC.md § 2. 改动 B — 落地惩罚改罚尾部、罚在触地前 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 A suspended (no-load) test acquits or convicts the actuator before you blame authority
suspended-test-isolates-actuator-authorityBefore attributing a failure to actuator authority, measure no-load tracking error and steady-state torque fraction; blame authority only if the task fails while the error grows with demanded force - and then fix gains or targets, not training.
Symptom
hip_roll sagged 0.21 rad on the ground and saturation questions loomed over the sidewalk plan - was the roll axis physically too weak (authority), or was something else limiting it?
Context
Before C4, the roll-authority question was settled by measurement triage: suspended test (--suspend, feet off ground) showed hip_roll tracking error 0.0008 rad - actuator acquitted; the entire 0.21 rad ground sag is load-induced. Steady-state torque was 25% of limit - 75% margin remains. Since sidewalk needs lateral force, not exact angles, authority was ruled "not a hard limit", with a pre-registered criterion for when it WOULD become one: sidewalk fails to track AND roll error keeps growing - then the fix is raising hip_roll kp or lowering the vy target, not more training.
Change
Hypothesis "roll authority insufficient" demoted from blocker to a monitored branch with an explicit trigger condition; C4 proceeded.
Outcome
Later open-loop probes confirmed the actuator could produce the behavior (sidewalk feed-forward ran at full amplitude, 5/5 survival), and the eventual C4 failure causes were measurement and reward, never authority.
Mechanism
Suspended vs loaded comparison separates the actuator's closed-loop competence from the load path: tiny no-load tracking error means the motor/controller is fine and any loaded deviation is statics (gravity / stiffness budget, kp trading error for force). Torque-fraction measurement then bounds how much force headroom actually remains.
Applies when
- suspecting an axis is "too weak" for a new skill
- large position sag on a loaded joint
- deciding between hardware fix, gain change, and more training
“吊挂（--suspend）实测 hip_roll 跟踪误差 0.0008 rad → 执行器无罪，地面下垂 0.21 rad 全是负载所致；稳态占限扭 25% → 仍有 75% 扭矩余量。… 判据：若 C4 出现「侧走跟不动且 roll 误差继续变大」，那才是权限账 … 解法是提 hip_roll 的 kp 或降 vy 目标，不是硬训。”
train/C_LADDER_RUN.md § 3d. roll 权限：已部分澄清，不是硬上限 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) When the eval proxy has a known systematic bias, gate on within-proxy differences, not absolutes
relative-metrics-survive-proxy-biasWhere the evaluator is known-biased, design gates as within-evaluator contrasts (left vs right, A vs B, pre vs post) that cancel common-mode error; reserve absolute thresholds for quantities whose proxy calibration has been checked.
Symptom
MuJoCo systematically overestimated yaw turn gain (1.58/2.45 where Isaac measured 1.04/1.02), and friction alignment recovered only part of the gap (mu 1.0 -> 0.6 pulled it to 1.43/2.27) - so any absolute turn-gain acceptance threshold in MuJoCo would be judging the proxy's bias, not the policy.
Context
The acceptance criterion was rewritten to use only the left/right difference of the turn gain (standard: <20%): both directions pass through the same biased proxy, so the bias largely cancels in the difference while the policy's chirality - the thing being gated - survives. The absolute-gain row was dropped: "转向只判左右差,不判绝对值". A parallel task was still opened to align MuJoCo contact parameters to Isaac's (mu, restitution), since drift proved highly friction-sensitive (straight-line drift -65 deg at mu 1.0 vs +1 deg at mu 0.4) - bias reduction and bias-robust metrics proceeded together.
Change
Gate metric changed from absolute turn gain to left-right gain difference; proxy-alignment work scheduled separately rather than blocking acceptance.
Outcome
Turn acceptance became meaningful across proxy versions (v5 70% -> v6 19% difference measured the real improvement) while the absolute bias question was pursued without holding the ladder hostage.
Mechanism
A systematic multiplicative or additive proxy bias applies to both arms of a mirrored measurement; differencing (or ratioing) mirrored conditions cancels the common-mode bias to first order, leaving the asymmetry signal. Metrics built this way remain valid while the proxy is imperfect - which it always is somewhere.
Applies when
- a sim proxy disagrees with the trainer or hardware in absolute terms
- writing acceptance thresholds for direction-paired skills
- proxy calibration work would otherwise block a ladder
“转向只判左右差,不判绝对值 —— MuJoCo 的偏航增益系统性高估(实测 1.58/2.45 vs Isaac 1.04/1.02),摩擦只能解释一部分(μ 1.0→0.6 仅拉到 1.43/2.27)。差值是相对量。”
train/WALK_V5_SPEC.md § 6. 验收 Cross-simulator gate (Isaac Lab to MuJoCo) comes before any hardware attempt
sim2sim-gate-before-sim2realGate every policy through a second simulator with an independently built plant before hardware; treat sim2sim failure as a contract or overfitting bug, and sim2real failure after a sim2sim pass as a plant/actuator gap.
Symptom
A policy that only ever ran in its training simulator carries untested dependencies on that simulator's solver, contact model, and defaults; the first place those dependencies surface should not be the real robot.
Context
Standing order of operations for the whole Lucen program, recorded as the opening line of the experience log: train in Isaac Lab, gate in MuJoCo, only then go to hardware. The MuJoCo side is the same plant used for evaluation batteries, so a sim2sim pass also validates the exported policy + contract (obs ordering, scales, defaults) outside the training stack.
Change
Pipeline rule adopted - every checkpoint must pass the MuJoCo evaluation battery (sim2sim) before it is considered for real deployment (sim2real).
Outcome
Used generation after generation as the cheap filter; hardware sessions only ever received policies that had already survived a second simulator.
Mechanism
Two simulators disagree exactly where a policy is overfit to simulator-specific artifacts (contact softness, integrator, default parameters, obs conventions); a cross-sim transfer catches contract bugs and solver overfitting at zero hardware risk, so real-robot failures that remain are attributable to genuine plant/actuator gaps.
Applies when
- planning the path from training to first hardware trial
- exported policy behaves differently outside the training framework
- triaging whether a real-robot failure is contract vs plant
“先sim2sim - 从isaaclab 到mujoco / 再sim2real”
Experience.md § opening lines (1-2) 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 A curriculum ramp keyed to the process step counter re-fires on every resume - and shipped policies never saw the penalty
curriculum-counter-lineage-stepsKey every curriculum/ramp schedule to lineage-cumulative progress, not per-process counters; and audit where your shipped checkpoints sit relative to every ramp - a penalty that no product ever experienced is not part of your training.
Symptom
vx+0.30 died at a fixed relative time in every resumed run: resume at 500 -> slide at 1100, zero at 1300; resume at 700 -> slide at 1300, zero at 1500 - absolute depths offset by exactly the resume offset, relative timetable identical.
Context
ramp_reward_weight (the saturation penalty ramp) read env.common_step_counter, which restarts at 0 for every run including --resume. So start_step=600 meant "600 iters after THIS resume", not "lineage iteration 600". The A/B arm pair was the clean proof: their env.yaml differed only in log_dir, only resume point distinguished them, and the omni CurriculumManager had exactly one active term - nothing else could produce that timetable. Second consequence: every shipped checkpoint (s1e-500 at +500, C2-700 at +200, A800 at +100) was selected before its run's +600, so the saturation penalty weight was 0.000 for every product ever shipped - explaining saturation 33% and raw |action| 1.9 against clip 1.0 (hip_roll in bang-bang), i.e. half the heat budget.
Change
Two independent recommendations recorded: (1) make the ramp count lineage steps (add the checkpoint's iteration offset on resume) or pin terminal weights in downstream rungs instead of ramping; (2) give the saturation penalty its own rung - never mixed into a skill-learning rung (that would be two variables again).
Outcome
Explained the recurring +600 death of the highest-amplitude command and the persistent actuator saturation of all shipped products with one root cause; honest caveat booked (at +800 the weight is only -0.086, small, but vx+0.30 is the command demanding the largest action amplitude, so it is squeezed first).
Mechanism
Resumable training splits "the lineage" from "the process"; any schedule keyed to process-local counters silently re-applies its transient to every descendant run, and any product-selection habit that picks checkpoints early systematically samples the pre-ramp regime - the curriculum exists in the config but never in any shipped policy.
Applies when
- resumed/forked training with any scheduled reward or DR ramp
- a metric dies at a fixed offset after each resume
- shipped policies show behavior a late-schedule penalty should prevent
“ramp_reward_weight 读的是 env.common_step_counter，它每个 run 从 0 开始，--resume 也不例外。… 原始 run / 臂B | 500 | 1100 = +600 | 1300 = +800；臂A | 700 | 1300 = +600 | 1500 = +800 … 所有出品其实从没见过饱和罚。… 这解释了 sat_max_pct 33%、raw |a| 最大 1.9（clip 是 1.0）—— hip_roll 一直在 bang-bang，而罚它的那一项权重恒 0。热账的一半在这里。”
train/C_LADDER_RUN.md § 3g. 系统性问题：saturation_ramp 每次 resume 归零 A DR tail the robot never has is pure cost - stage deterministic plant levels instead of one wide uniform
dr-tail-plant-continuationSet every DR range from the measured deployment distribution and cut tails that hardware cannot produce; when an axis changes the controller's character (delay, major gain regimes), prefer staged deterministic levels with gates over one wide uniform.
Symptom
Two consecutive lineages (s1e, s1f) trained under uniform latency DR (0, 0.06 s = 0-3 frames) both converged to drag-glide gaits - buying survival under heavy delay by giving up swing (3.6 mm) - even though the real pipeline never exceeds ~2 frames.
Context
The account: roughly 1/3 of training quality was spent on the >2 frame tail that hardware never presents ("uniform 尾部 ~1/3 训练质量 花在真机不出现的 >2 帧上"). The deeper reading came from the user: uniform 0-3 frames is not merely tail-heavy - it "把性质不同的控制系统 混进同一 PPO batch" (mixes qualitatively different control systems into one PPO batch); a 0-frame and a 3-frame plant demand different controllers, and one policy trained on the mixture serves neither. The S2 v2 ladder therefore redefined latency "从「随机化参数」重新定 义为 actuator/control plant 的一部分": deterministic FIFO levels, staged 1 frame then 2 frames (lo=hi so fractional interpolation degenerates to exact N frames, synonymous with the harness --delay N), each level gated by the fixed acceptance battery - a plant continuation, not a randomization.
Change
Latency DR replaced by staged deterministic levels covering the measured 1-2 tick reality with no tail; each stage a separate continuation rung with the standard gate and rollback.
Outcome
The s2_lag1 rung showed the clean-signal benefit immediately (survival 20/20, heading 6x recovery) with the swing cost booked honestly (21 -> 12 mm, half-pass, ladder paused for adjudication); the drag-glide attractor from uniform tails did not recur.
Mechanism
DR asks one policy to cover a plant family; when part of the family is fictitious, the policy pays real capability for fictitious robustness, and when family members demand structurally different controllers, gradient averaging produces a compromise controller optimal for none. A measured, discrete plant set matches the actual deployment support and keeps each rung's training signal coherent.
Applies when
- policies converge to degenerate gaits that buy worst-case survival
- a DR range extends well past the measured hardware range
- choosing between wide randomization and a staged ladder on an axis
“两轮实证(s1e/s1f)宽尾延迟 DR 逼出拖地滑行 … uniform 0~3 帧不止尾重,而是把性质不同的 控制系统混进同一 PPO batch;1→2 帧确定性分级 = plant continuation,训练信号干净得多—— latency 从「随机化参数」重新定义为 actuator/control plant 的一部分。”
train/OMNI_V0_SPEC.md § 4. v2 阶梯 (2026-08-07 用户定) 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 A soft joint-limit penalty charged the standing pose itself - the geometric-zero knee sat on its hard limit, so stand_v1 bent its knees to dodge 0.419 per step and leaned 4.1 deg forward; excluding the knee gave 0.24 deg
soft-limit-penalty-charges-nominal-poseBefore training, evaluate every penalty at the nominal pose; if a joint's soft limit sits inside the pose the task requires (a straight knee on its hard stop), exclude that joint from the soft-limit penalty and let the action clip enforce the hard limit.
Symptom
stand_v1 (retrained after the default pose moved to the CAD geometric zero and mirror augmentation was added) fixed left/right asymmetry (6.8 -> 0.0 deg) but settled at a 4.1 deg forward lean, where pure PD at the same default settled at 0.1 deg - the policy was actively pushing itself forward, which is exactly the real robot's failure direction.
Context
soft_joint_pos_limit_factor = 0.9 shrank the knee's soft limit to -/+0.1047 rad, while the geometric-zero default has the knee at q = 0, exactly on the hard limit. Standing in the nominal pose therefore paid 0.2094 x 2.0 = 0.419 per step in dof_pos_limits (alive earned only 0.5). The policy's way out was to bend the knees to -/+0.1013 rad, and the cost was the forward lean.
Change
stand_v1b: the knees excluded from dof_pos_limits (a straight knee IS the standing pose; the hard limit is still enforced by the action clip). No other change.
Outcome
stand_v1b: max tilt 0.3 deg, steady tilt 0.24 deg, asymmetry 0.1 deg, height 0.384 m - exactly nominal - with knees at -0.0007 / +0.0005 rad. It became the standing release used on the real robot, and later the standing side of the recovery switch. The same exclusion was carried into the recovery contract (knee at the clip in the standing pose) and the one-leg reward table.
Mechanism
A limit penalty whose soft boundary lies inside the nominal pose turns the nominal into a taxed state, and the policy buys its way out with whatever posture change is cheapest - here a knee bend paid for with lean.
Applies when
- a standing or default pose has a joint at or near its hard limit
- a policy settles in a small steady tilt that pure PD does not show
- soft-limit factors shrink limits uniformly across joints
“`soft_joint_pos_limit_factor=0.9` 把膝软限位内缩到 ∓0.1047,而几何零位 default **膝盖 q=0 正好压在硬限位上** ⇒ 站在标称姿态每步白扣 `0.2094 × 2.0 = 0.419` (alive 才 +0.5)。策略只能屈膝到 ∓0.1013 躲罚,代价是躯干前倾 —— 恰好是真机的 失效方向。 … 修掉"软限位罚标称姿态"后重训(`dof_pos_limits` 排除膝盖)。**前倾问题彻底消失** … 不再屈膝躲惩罚,高度正好落回标称 0.3840。”
train/README.md § 三期: 镜像对称增强 + 站立 v1 (2026-07-28) / stand_v1b (2026-07-28): 站立定版 Three same-shaped judging errors - task metrics (survival, tracking, displacement) cannot stand in for posture metrics
task-metrics-vs-posture-metricsKeep validated posture-class rows (tilt, per-joint L/R asymmetry, temperature) in every acceptance battery alongside task rows; when operator feel contradicts the gates, suspect the metric class before the operator - and never build a new skill on what is actually an asymmetry defect.
Symptom
The C2 product judged "full pass" on task metrics (A800: turn-gap 7 pp, vx+0.30 19/20) felt WORSE in the operator's hands than the half-pass 700: A800 tilted up to 12.90 deg (700: 6.64), drifted left while standing, showed larger per-joint asymmetries, and ran its hip_rolls 5 degC hotter.
Context
The re-judgment catalogued three same-type metric errors in one campaign: (1) stand judged by SURVIVAL - missed 0.5-1.4 m wandering; (2) stand ranked by DISPLACEMENT - ordering was opposite to real feel (tilt ordering matched); (3) chirality judged by wz-tracking GAP - measured turning symmetry while the robot's actual disease was postural left/right asymmetry, "两个不同的东西,且结论相反". Common pattern named: "我一直用「任务指标」当判据,而真机手感对应的是「姿态 指标」… 任务类指标不能替代它". The fix was already in the data: the per-joint left/right asymmetry table (printed identically by sim2sim and deploy) agreed with hardware in direction on every row - "判据可用、有预测力,我只是没把它写进 PASS 条件". Shipping decision followed the posture read: product reverted to 700 ("又一次「买到 精度、卖掉别的」"), and the C4 root moved to 700 as well, with the sharpest line of the episode: A800's left-drift "like sidewalking" is probably its frontal-plane asymmetry defect, not a capability - "在缺陷上建能力是危险的".
Change
Two posture quantities with demonstrated real-robot predictive power promoted into every PASS battery: tilt-max median and per-joint left/right asymmetry (both sim-computable, deploy-homologous); motor-temperature readout added to session close-out.
Outcome
Deployment flipped to the posture-better checkpoint; the hip_roll temperature table (43-48 degC vs 25-28) confirmed the earlier 90%-of-heat account; the run-line acceptance battery inherited the posture rows from birth ("任务类替代不了姿态类").
Mechanism
Task metrics measure goal attainment under the evaluator's episode definition; posture metrics measure the body state trajectory that operators, motors, and downstream skills actually experience. The two can rank candidates oppositely because task success tolerates postural pathology - so a battery without posture rows is blind to exactly what hardware feel reports first.
Applies when
- hardware feel disagrees with a green acceptance table
- choosing between checkpoints that split task vs posture metrics
- selecting the root for a skill that resembles an existing defect
“共同模式:我一直用「任务指标」(存活 / 跟踪率 / 位移)当判据,而真机手感对应的是「姿态指标」(倾角、逐关节左右不对称)。→ 验收判据里必须有姿态类指标,任务类指标不能替代它。… A800 的「左飘像 side walk」很可能 … 是它更大的额平面不对称的表现 —— 在缺陷上建能力是危险的。”
train/README.md § C2 选点改判 (2026-08-09): 手性判据第三次选错指标 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) Order hardware runs by sim risk, gate each stage on the last, and put the fragile cell last with a spotter
risk-ordered-real-deploymentScript hardware sessions as a risk ladder: baseline first, sim-riskiest last with a spotter, suspended smoke before ground, each stage gated on the previous, environment (floor mu) recorded as a selection input - and stop at the stage that misbehaves.
Symptom
Five policy-x-gain combinations had to go on hardware in one session, with sim survival ranging from 20/20 down to 17/20 (and zero-command survival down to 2/20) - an unordered session risks breaking the robot on an avoidable run.
Context
The execution sheet fixed the order as sim-risk low to high, control baseline first (current SOTA establishes the floor reference), the fragile cell (fric-2400@kd1.0) last with a person spotting throughout. Stage gating: suspended smoke (feet off ground, 10 s each, all five pass before anything touches down) -> suspended with IMU and forward command (gait forms in the air) -> grounded runs -> speed raise only for combos that survived the previous stage -> zero-command tests only with a spotter, ordered by sim zero-cmd survival, with the 2/20 cell skipped by default. Preconditions include recording the floor material and estimating mu (if mu <~0.6, sim says pick the kd1.2 gain as main), port/CAN self-check, calibration frozen. Any stage failing stops the session at that stage: "任一段出问题就停在那一段, 不要跳到下一段".
Change
Session structured as a risk ladder with per-stage gates instead of a flat checklist; per-combo sim survival numbers written into the run table as the ordering key.
Outcome
The session design localized any failure to the cheapest stage that could reveal it, kept the robot safe for the informative fragile run, and made the control baseline available before any comparison run.
Mechanism
Hardware sessions consume a shared budget (robot integrity, battery, floor time); ordering by predicted risk means information is bought cheapest-first, and stage gates convert an expensive failure into a cheap earlier one. Baselines run first because every later reading is relative to them.
Applies when
- taking multiple policies/configs to hardware in one session
- a candidate is known-fragile in sim but must be measured
- writing a deployment runbook for a new robot
“跑序 = sim 风险从低到高, 最险的放最后 (依据 = 存活门/零指令存活) … ⑤ 是 sim 里最脆的一格 … 放最后跑, 全程留人扶, 起步即给 cmd, 零指令不做。… 任一段出问题就停在那一段, 不要跳到下一段。”
train/REAL_RUN_S2.md § 上机名单 / 全部命令 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 The hip_roll (l+r) asymmetry scalar predicted real-robot lateral drift - promote validated sim scalars into the gate
hip-roll-sum-predicts-lateral-driftHunt for cheap sim scalars that predict real-robot behaviors, validate them on direction AND ordering across multiple policies, then promote them into the acceptance battery; treat later violations as debt to justify in writing, not noise to ignore.
Symptom
A persistent hip_roll left/right asymmetry row in the sim2sim symmetry table had been dismissed as "calibration or mechanical asymmetry" noise; meanwhile real deployments drifted sideways by policy-dependent amounts.
Context
Forward-kinematics analysis reframed the scalar: both hip_rolls move the feet in +y for positive angle, so a same-signed (l+r) sum IS a lateral translation mode - the scalar is a direct lateral-drift bias estimate. Checked against real deployments: s1e (l+r = -0.0178, smallest magnitude) was the steadiest with least drift; 700 (+0.0253) drifted mildly left; A800 (+0.0267) drifted clearly left with the largest tilt 12.9 deg. Direction correct 3/3, ordering correct 3/3 (the log's heading calls it "四枚四中", four-for-four).
Change
The scalar was promoted into the acceptance battery as a posture-class criterion alongside tilt-max median: "hip_roll 左右不对称 |l+r| 不得比父代大" - doubling as a heat proxy (error ~ torque ~ heating).
Outcome
Used at every later gate; when the C4 product exceeded it by +0.005 rad (~+0.3 deg vs parent), the criterion was not silently waived - it was booked as explicit debt with a mechanism argument (the increment is task-required, far smaller than the sidewalk amplitude +/-2.2 deg) plus a related account (stand saturation 32.4% -> 37.2%).
Mechanism
A policy's static joint-angle bias in a translation-producing mode integrates into real-world drift; sim can measure that bias precisely and cheaply. A sim scalar earns gate status exactly when its predictions are validated against hardware in both direction and ordering - and a validated gate may only be exceeded with a written mechanism-level justification, never silently.
Conflicts
The log's heading says "四枚四中" (4/4) but the evidence table lists three policies and the text says "方向 3/3、排序 3/3"; the fourth instance is not shown in this file.
Applies when
- a real robot drifts or leans in a policy-dependent way
- deciding which sim measurements deserve gate status
- a validated gate criterion is marginally exceeded by a new product
“s1e | −0.0178（绝对值最小）| 微右、最不飘 | 三者中最稳、飘最小 ✓ … A800 | +0.0267 | 左、最飘 | 明显左飘、倾角最大 12.9° ✓ 方向 3/3、排序 3/3。 → 正式纳入验收表（与「倾角 max 中位」并列为姿态类判据）。”
train/C_LADDER_RUN.md § 3e. 顺带：hip_roll 左右不对称 (l+r) 就是横移偏置 —— 四枚四中 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