Auditing Dynamic Feasibility of Retargeted Humanoid Motion Data

Companion note draft v0.2 (2026-08-20; v0.1 archived in git history). Target: arXiv + workshop, submittable Sept 5. 4–6 pages. Status labels: sealed ✓ / sealed ✗ (kept) / measured / exploratory / pending 🕐 appear on every claim, including captions. Every number cites its artifact path (paper/RESULTS_LOG.md is the master index). Engine pins (used throughout): MuJoCo 3.11.0 (C) · MuJoCo Warp 3.11.0 · mjlab v1.6.0 · Newton commit 7bb6d02d (pip 1.6.0.dev0) · warp-lang 1.16.0 · Unitree G1 model g1.xml sha febdcbef… (plan/PREREGISTRATION_G1_clip44.md §Pins). The conformance certificate is a property of this version pair — see §6 and the re-certification spec.

Abstract

Large humanoid motion-tracking systems train on tens of thousands of retargeted mocap clips that pass kinematic quality control: joint limits respected, velocities smooth, feet roughly grounded. We show that kinematic QC misses a distinct and consequential failure class — references that are dynamically impossible for the target robot given the contacts they make available — and give a screen that catches it at ~1 CPU-second per clip: contact-free inverse dynamics for the base wrench the environment must supply, then a torque-limited linear program over friction cones at the contacts within reach. On a 10,705-clip AMASS→Unitree-G1 bank [measured], 22.8 % of clips demand unsupported forces exceeding half the robot's weight for more than 10 % of their frames; the rate varies from 0.1 % to 100 % across source datasets under a single retargeting pipeline, marking it as a corpus-and-pipeline property rather than a motion property. A second production bank — the 4,950-clip BONES-SEED corpus that trains SONIC, screened by an independent re-implementation on a pre-registered test — returns 0.14 % [measured], which settles the reading: prevalence belongs to a particular corpus-and-pipeline pairing, must be measured per corpus rather than assumed from another, and at ≤ 1 CPU-second per clip is cheap enough to be a standing release gate. We validate the screen on the clip that broke a failure-adaptive training curriculum: a kneel whose retargeted descent is airborne for a full second (median 329 N unsupported against a 327 N robot weight) [measured], which a pre-registered physics-parameter gate could not rescue in any of ten configurations [sealed ✗, kept]. A two-seed rollout anomaly — increasing motor strength worsens tracking inside the flagged window — does not generalize: its sealed family/control test fails all three criteria [sealed ✗, kept]. We quantify downstream costs: contaminated evaluation sets (29/100 held-out clips) [measured], poisoned failure-weighted samplers [sealed ✓], and difficulty labels that transfer across policies only once feasibility features are added (Spearman 0.567 → 0.609 on the headline pair, permutation p = 0.010; direction positive on 6/6 pairs, 4/6 significant) [sealed ✓]. We release the screen, an error-class checklist for coupling physics stacks to RL harnesses, and recommend feasibility-stratified evaluation.

1. The failure class kinematic QC cannot see

A 100-clip Unitree-G1 tracking curriculum of the BeyondMimic/mjlab family collapsed: its failure-adaptive sampler concentrated 87–89 % of exposure on one kneel-and-crawl clip in 3/3 seeds and lost to uniform sampling (held-out survival 0.780 vs 0.810; per-seed Δ +0.030/+0.028/+0.030) [sealed ✓; reports/campaign_summary_3arm.json, reports/A5_coverage_dose.json; sampler mechanism filed as mjlab #1153 / whole_body_tracking #73]. The clip itself passed every kinematic check: zero joint-limit violations against the G1's ranges, joint speeds ≤ 5.6 rad/s, smooth [measured; plan/N1_RESULT.md]. A pre-registered gate over physics parameters — action delay, motor strength, foot friction, contact stiffness, contact model, torso CoM; 1,440 paired worlds — failed its own sealed criteria: no intervention changed the clip's survival in any configuration [sealed ✗, kept; plan/PREREGISTRATION_G1_clip44.md 41e4b20c, plan/G1_RESULT.md, reports/G1/run0/].

What no kinematic feature could see: for the second between 0.75 s and 1.75 s of the retargeted descent, no part of the robot's collision geometry is within 6 cm of the floor — the feet float 7–10 cm up while the pelvis falls 0.79 → 0.40 m — so a median 329 N (p90 348 N) must come from nowhere; the robot weighs 327 N [measured; reports/N1_clip44_knee_id.json]. The mechanism is prosaic: the human sat back onto their heels; the G1's legs cannot fold that far; the retargeter resolved the conflict by lifting the legs instead of lowering the root. Under frame-0 starts the clip's survival is 0.00 for every policy we trained; the previously reported 0.31 was an artifact of averaging over start offsets that skip the impossible segment [measured; reports/N3_baseline_uniform-s1_strat.csv] — one of two evaluation-protocol lessons this note carries (§5). Figure 1 shows the anatomy.

Figure 1 [measured]: reference stick-frames (red panels: no contact available) and the torque-limited unsupported force; descent and rise pinned at ≈ body weight. Script + data: tools/n1_knee_id.py → reports/N1_clip44_knee_id.json; rendering reports/upstream_drafts/clip44_airborne_repro.png.

2. The screen

Per frame of a retargeted reference: (i) q, q̇ from the clip; q̈ by smoothed central differences; (ii) contact-free inverse dynamics (mj_inverse, contacts disabled) → the 6-D base wrench W the environment must supply, plus contact-free joint torques; (iii) candidate contacts = collision geoms within gap = 6 cm of the plane (exact geom distances); (iv) two solves over pyramidal friction-cone generators at those contacts: NNLS for the unconstrained residual, and a torque-limited LP (cone coefficients + L1 wrench slack, joint torques constrained to actuator force ranges) for the smallest unsupported wrench any controller could leave. Clip features: airborne_frac (no candidate contact), infeasible_frac (torque-limited unsupported > ½ weight), unsupported_impulse_per_weight_s (seconds of free-fall equivalent), torque_infeasible_frac, each under a uniform-μ and the simulator's own per-geom contact model. ~1 CPU-second per clip.

Stated limits: plane-only terrain; true ballistic flight is exempt by construction (a body in free fall demands no support — q̈ ≈ g leaves no unsupported wrench), though non-ballistic floating around take-offs is flagged, correctly (§4); the screen certifies the retargeted output on the target robot, never the source mocap. The gap choice matters and is justified rather than assumed [measured; reports/N1_gap_sensitivity.json]: at 3 cm even the feasible control is flagged 42 % (the bank carries a systematic ~3 cm stance-clearance offset from retarget ground alignment, so real contacts sit just outside a 3 cm band); at 10 cm the screen degenerates the other way (geometry 7–10 cm airborne is granted as a contact candidate and the attractor's descent reads 0 % infeasible). 6 cm sits between the two failure modes — above the bank's clearance offset, below physically bridgeable distance. The ½-weight bound is not delicate: the attractor's flagged-frame fraction moves only 15.1 % → 13.1 % → 12.5 % across bounds of 0.25/0.5/0.75× weight, because the flagged mass concentrates near 1× weight. Tool: refeas v0.1.0 (Apache-2.0; MuJoCo + SciPy; G1 worked example whose synthetic hover clip is flagged at 45 % infeasible — refeas/examples/demo_hover_brief.json).

3. Validation on the clip that mattered

Phase-resolved verdict for the attractor clip [measured; reports/N1_clip44_knee_id.json]: standing 0–0.75 s supported (torque-limited residual 0 N); descent 0.75–1.75 s airborne in 86 % of frames, median unsupported 329 N; kneel/crawl 1.75–7.25 s fully supportable within actuator limits even with the simulator's frictionless knees; rise 8.0–8.5 s airborne again; final stand supported. A matched-easy control from the same bank is supported at every frame (torque ratio p95 0.66) [measured; reports/N1_CMU76_knee_id.json]. The family is systematic: 20 of the clip's 40 nearest kneel/crawl neighbours exceed the 10 %-infeasible-frame threshold [measured; reports/N3_candidate_feasibility.json].

Independent corroboration, from rollouts that never see the screen: with a calibrated paired-rollout statistic (signed replicate-mean effects over a published identical-physics floor of ≈ 0 ± 1 mm — the floor matters because identical physics already diverges 2.5–8 mm within seconds; reports/G1/run0/g1_v2_summary.json), +15 % motor strength worsens body-position tracking by +15.0/+16.0 mm exactly inside the screen-flagged airborne window and ≈ 0 mm (+0.3/−1.4) in the supported standing window, replicated across two IC-seed sets (instrument-wide agreement across all 36 axis×clip effects: r = 0.92) [exploratory — two cases; plan/N5_RESULT.md, reports/G1/run1_seed1/g1_v2_summary.json]. The sealed generality test on 12 family clips versus 12 feasible controls fails: 7/12 family clips pass the +5 mm criterion (needed 8), only 4/12 controls stay within 2 mm (needed 8), and just 2/7 positive cases are airborne-localised [sealed ✗, kept; reports/P_SIGN/run0/p_sign_summary.json, plan/P_SIGN_RESULT.md]. The anomaly therefore supplies no rollout-only detector claim.

4. Prevalence

Bank-wide screen — AMASS→Unitree-G1, this pipeline, this robot (10,705 clips) [measured; reports/feasibility_all/prevalence_report.txt, sentinel COMPLETED]:

By source: GRAB 0.1 %, TCD 1.6 %, KIT 16.5 %, BMLmovi 21.8 %, Eyes-Japan 23.2 %, BMLhandball 27.4 %, CMU 39.7 %, ACCAD 41.7 %, HUMAN4D 55.4 %, Transitions 89.6 %, CNRS 100 % — three orders of magnitude of spread under one retargeter and one robot. Hand-checks of the extremes [measured; reports/upstream_drafts/CNRS_AUDIT.md]: the CNRS clips are ordinary fast walks in all five severity strata inspected: median lowest-geometry clearance spans 5.1–9.7 cm and each develops a 1.34–3.76 s floating run — an ingest/retarget root-height defect, not exotic content. Transitions is mixed rather than uniformly acrobatic: among five severity-stratified clips, the air-kick and 360° jump/twist are content-driven, but two locomotion clips and a maximum-severity standing punch are ingest defects (3/5 ingest, 2/5 content, 0/5 scene mismatch). The dynamic category's rate is therefore partly content-inflated and partly output-side; the ground category's 39 % cannot be explained by acrobatics and is the attractor's family writ large. Figure 2 [measured]: paper/figures/f4_prevalence.png (script f4_prevalence.py, data reports/feasibility_all/feasibility.csv).

A second bank, a second pipeline [measured; pre-registered]

The 22.8 % above is a measurement of one pipeline, and reporting it alone would invite the inference that retargeted humanoid banks are generically ~20 % broken. They are not. The same method, re-implemented independently against a different G1 model file (g1_29dof_rev_1_0.xml, sha 15a330f1…; μ 0.7 rather than 0.6, same 6 cm gap, same ½-weight bound), was run over every clip of the BONES-SEED bank consumed by SONIC:

The measurement was registered before it was run, with its consequence pre-committed — a rate under 10 % descopes a planned feasibility-hygiene training ablation on that stack — and the ablation has been descoped (P10, GR00T-WholeBodyControl/docs/prediction_register.md; screen gear_sonic/research/hygiene/screen.py; 4,950 clips in 131.7 s wall on 8 CPU workers, 0 failures, 0.145 CPU-s/clip = 0.84 ms per screened frame).

Two things must be said about it, in this order. First, the class is not absent from the cleaner bank: seven clips exceed the threshold, five of them jumps — four named for the 50 cm box they jump onto or off, which is absent from the flat scene, plus a high jump; the remaining two are a kick-back (kick_back_001, 0.47 infeasible) and a burpee. These passed both kinematic QC and a shipped release filter, and they are a real data defect, though a different one (a scene/reference mismatch whose fix is terrain or exclusion; root projection is the wrong operator, §8). Second, and this is the finding: prevalence is a property of a particular corpus and pipeline, not of the practice of retargeting, so it has to be measured per corpus and can be, in minutes. That is what turns the screen from a one-off audit into a release gate. The comparison bounds generality rather than isolating a cause — two source corpora, two implementations of one method, a release filter on one side only — and the controlled version (two retargeters over the same source clips) is not run.

A same-clip implementation check removes one part of that caveat [measured; reports/feasibility_xcheck/]. We applied both implementations to a deterministic stratified 20-clip sample from each bank (all seven BONES-SEED flags + 13 sampled feasible clips; ten AMASS flags including #44 + ten sampled feasible clips). Across the 40 clips, CLIMB and SONIC scores have Spearman ρ = 0.984 for infeasible_frac and 0.997 for airborne_frac; the strict > 0.10 verdict agrees on 39/40 clips (97.5 %, Cohen's κ = 0.948; matrix: 16 both flag, 23 neither, one SONIC-only, zero CLIMB-only). This closes implementation agreement on the stratified panel, not the corpus/filter confound and not a prevalence estimate; the full table is Appendix A3.

The same run is also the clearest evidence for keeping airborne_frac and infeasible_frac as separate axes rather than collapsing them into one "grounded" flag: seven kneeling_loop_* clips sit at airborne fraction 1.000 with infeasible fraction 0.000 — feet 7–9 cm off the floor for the entire clip, weight carried on the knees, supportable at every frame. A filter reading "airborne" as "broken" would delete exactly the rare ground-contact behaviour these banks are short of.

5. Downstream costs

Samplers [sealed ✓]: a failure-weighted curriculum reads "impossible" as "maximally informative" (§1). Evaluation [measured]: 29/100 of our held-out clips are flagged; flagged clips score 6.0–8.4 points below each policy's all-clips aggregate (8.4–11.8 below its feasible stratum; reports/N_atlas_v21.json endpoints_2b) and cannot separate training methods — the grounded-vs-uniform endpoint edge is +0.025 on feasible clips and −0.009 on infeasible ones (reports/N_atlas_v21.json endpoints_2b). Our project now reports feasibility-stratified endpoints under a sealed policy whose threshold predates it (plan/GLOBAL_EVAL_ADDENDUM.md a93a87a0). Difficulty labels [sealed ✓ / sealed null kept]: reference-feature difficulty models transfer across policies at ρ = 0.567–0.579; adding the three screen features moves all six transfer pairs in the right direction, four significantly beyond a 200-draw random-feature baseline (headline 0.567 → 0.609, p = 0.010) (reports/N_atlas_v21.json); bank-relative support features alone did not clear the same baseline — a sealed null we keep (reports/N2_atlas_support.json). A bounded near-miss [sealed null, kept]: a second reference artifact — hand–hip interpenetration > 1 cm on a median 13 % of frames, taxing the self-collision reward on accurate tracking — does not predict difficulty beyond the feasibility flag (heldout partial ρ −0.04…−0.15, no positive CI on any arm, sealed rule 0/3; plan/P_TAX_RESULT.md, reports/P_TAX_result.json). It remains a hygiene recommendation, not a claim.

6. The harness had to be proven first (and the proof is version-pinned)

None of §3's rollout numbers would be meaningful over a broken coupling. Before any of them, the same engine (MuJoCo Warp 3.11.0) reached through two integration stacks — mjlab v1.6.0 directly vs Newton 7bb6d02d via SolverMuJoCo, with classic MuJoCo 3.11.0 (C) as third referee — was driven to per-substep agreement |Δq̇| ≤ 3×10⁻⁵ rad/s, after a four-stage elimination surfaced four silent coupling errors whose combined effect (a 40-point survival fork with matching mean error) we had initially misread as a physics finding and formally withdrew [measured + withdrawn verdict; plan/S1_RESULT.md, reports/S1_KIT1226_n32_absorb.json]. The generalised checklist — nine error classes, each with symptom, detector, and instance — is this note's Appendix A and ships in the tool's docs. The certificate is a property of the pinned version pair (release numbering of the Newton line differs from this checkout's package metadata; the commit hash is the pin); a re-certification against Newton 1.0 GA (which changed the collision stack) is specified as future work [pending 🕐; plan/SPEC_newton_recert.md].

7. Recommendations

1. Screen before training (~1 CPU-s/clip) and measure prevalence on your own corpus — the two banks screened here differ by a factor of 160 (§4), so no published rate transfers; publish per-clip feasibility flags with datasets and re-run the screen as a release gate on every new corpus.
2. Report feasibility-stratified endpoints; never average survival over start offsets that straddle infeasible segments.
3. Retargeting pipelines: resolve unreachable postures by root-height/contact projection, not limb lifting; add a per-source stance-contact consistency pass (the CNRS class); audit reward terms against the reference itself.
4. Couplings: certify any second physics stack at the substep level against the Appendix-A checklist before believing closed-loop differences; pin and publish both stack versions.
5. For flagged-but-wanted motions, prefer repair or segment eligibility over blind deletion: clip pruning at 12.4% contamination produced a sealed null with feasible heldout Δ −0.0101, while segment curation retains feasible material. Repair-all produced a positive deployed- reference contrast (+0.0397) but missed its +0.05 benefit and coverage gates; raw-reference policy transfer was −0.0036 [sealed ✗ + measured, plan/E_HYG_RESULT.md, plan/N7_RESULT.md].

8. Deliverables and economics

The economics, stated once [measured]: the screen costs ~1 CPU-second per clip — ~3 CPU-hours for this entire bank — and the geometric repair ~3 CPU-seconds per recoverable clip, against training runs of 10³–10⁴ GPU-hours (SONIC-scale: 21,000 GPU-hours). Feasibility hygiene is four to five orders of magnitude cheaper than the training it protects, and the exposure it reclaims is not marginal: a failure-weighted sampler concentrated a mean 48.8 % of its draws on a single clip, at least 21.9 % of them on the impossible one [measured; reports/wasted_exposure_accounting.json].

Three deliverables, at three pipeline stages: (i) refeas v0.1.0 — the offline pre-training screen (Apache-2.0, github.com/linjiw/refeas; version hash pinned in the sealed evaluation policy). (ii) contact-projection repair (tools/repair_contact_projection.py) — recovers the root-floating defect class in ~3 s/clip [measured: the attractor 0.13 → 0.00 at 8.2 cm; a 100 %-flagged subset's walks 0.66 → 0.01] while refusing genuine ballistics via an over-repair budget. Census over every flagged clip [measured; reports/repair_census/summary.md, sentinel present; membership correction C4]: 65.8 % of the strict 2,442 flagged clips (1,606) are auto-recoverable under the legacy 15 cm budget (the historical 2,443-row directory additionally contains one feasible no-op control; 73.1 % of its 10–25 % band, 61.7 % of its > 25 % band; ground category 68.1 %, quiet 80.9 %, dynamic 51.2 % — the last correctly depressed by refused ballistics). The practical split on this bank: roughly two-thirds of the contamination is a 3-second script; one-third needs upstream re-retargeting or higher-order (IK/time-warp) repair. That split is no more portable than the prevalence: where the defect class is scene mismatch rather than root float — the second bank's box jumps (§4) — this operator is the wrong one and the recoverable fraction would be near zero. (iii) evaluation & monitoring protocols — stratified-start evaluation, feasibility-stratified endpoints (sealed policy a93a87a0), the Appendix-A coupling checklist. The proposed rollout-only infeasibility detector is explicitly not a deliverable: P-SIGN rejects it [sealed ✗, plan/P_SIGN_RESULT.md].

Exact qualification changes the claim [unsealed measured implementation; not policy-validated]. The 65.8% census routes legacy root-only repairs; it does not promote them to training. DFRP v1 instead freezes a stratified 26-flagged + 4-control panel and admits a repair only when residual infeasibility is ≤ 5%, root displacement is ≤ 8 cm, joint limits pass, contact-IK residual is ≤ 10 mm, legal 50-step starts exist, and every source/repair/sidecar/unit-table identity is hash-bound. 22/26 flagged candidates pass and 4/4 controls remain byte-identical. The curated 26-clip view exposes 36 exact units and 10,561 legal starts; two residual-feasibility and two IK-qualification failures remain excluded (reports/dfrp_v1_exact_panel/iter1/result.json, plan/DFRP_V1_EXACT_PANEL_RESULT_2026-08-21.md). This is an implementation gate on a panel, not a bank-wide recovery rate or a policy-benefit result. The separate residual and IK failures are the reason those gates may not be collapsed into one "repaired" label.

Segment-level curation, and why its value is framework-dependent [measured; reports/segments_tier800/segments_guard0.csv, …_guard1.0.csv, reducer tools/screen_segments.py]. Pruning and repair are not the only options: a flagged clip is usually mostly feasible. Re-screening the 99 flagged clips of an 800-clip training tier at segment resolution (99 clips in 45 s wall on 6 nice'd CPU workers) — contiguous severe windows, guard-band expansion, minimum-length filtering, projection onto the sampler's bin grid — gives:

Against the whole tier (152.4 min), clip-level pruning discards 13.3 % of its duration while segment curation hands back 8.2 % at guard 0 s and 3.8 % at guard 1.0 s. Identical screen, identical clips: the value of segment-level curation falls as the policy's reference lookahead grows — the guard band is a property of the training framework, not of the data. Clip-level pruning is therefore a lower bound on what feasibility hygiene can buy. Caveats carried: the 1.0 s minimum segment length and the strict bin-eligibility rule (any severe frame disqualifies a bin) are choices, not measurements, and these are duration recoveries — no policy has yet been trained on curated segments. The related clip-pruning training arm is a sealed null: feasible heldout survival 0.918→0.907 (Δ −0.0101, one-sided permutation p=0.951), with zero-shot ground Δ −0.0354 inside its pre-registered coverage-cost bracket (reports/E_HYG_result.json, plan/E_HYG_RESULT.md). This makes segment-level eligibility, which keeps feasible portions, the sharper next training test; it does not turn the duration measurement into a performance claim.

The existing exact-segment wiring pilot does not yet supply that policy test [exploratory]. Its runtime is mechanically clean—exact truncation and zero invalid/censored trials—but conditional failure saturates and leaves the adaptive distribution only 0.014 TV from its capped uniform control (reports/segment_v2_pilot/result.json). It is therefore a failed manipulation check, not evidence for or against adaptive segment allocation. A new arm must pass a predeclared allocation band before its outcome is interpreted.

Appendix A — coupling-error taxonomy

(→ appendix_coupling_taxonomy.md, shipped verbatim; also refeas/docs/COUPLING_TAXONOMY.md.)

Completed sealed work retained here: N3 af1b7c9f mixed outcome; E-HYG a5494b7c null; P-SIGN c7916e8c fail; soft FGAS 3521c80e implementation-gate fail; N7 repair 90da8a08 joint fail. The segment-v2 pilot is exploratory and fails its allocation manipulation (TV 0.014), so adaptive exact-segment sampling remains untested. Pending work: E3 support moderation 2c38845b 🕐 and a separately sealed segment-native follow-up under the v2 lifecycle/evaluator. Figures: F1 anatomy, F2 prevalence — scripts + data in paper/RESULTS_LOG.md. Author list / acknowledgements TBD with Linji. Upstream note drafts are separate documents awaiting approval (reports/upstream_drafts/).

5b. What tracking an infeasible reference does — measured in sim, predicted for hardware

(Added v0.3, 2026-08-20; requires review before submission — logged in REVIEW addendum.)

Measured [reports/effort_sat_at_fall.json, from reports/G1/run0/armA.npz]: tracking the airborne descent, zero of 29 actuators saturate at any point in the supported phase; within 0.6 s of losing foot support, ≥ 4/29 actuators pin at ≥ 98 % of force range in 8/8 replicates — exactly 5/29 in 7/8, mean 16.8 %. The failure mode is not gradual degradation — it is a commanded hover ending in an unplanned ~0.3 m fall onto joints that are not landing gear (estimated ~2.4 m/s touchdown, ~95 J to dissipate [estimate]). Predicted hardware phenomenology [exploratory, sim-grounded, no hardware claims]: impact loading on wrists/knees at every attempt of the segment; current/thermal-limit bursts invisible in average torque. The proposed gain-response diagnostic does not survive its sealed test: P-SIGN finds the predicted sign in 7/12 family clips, only 4/12 clean controls, and airborne localisation in 2/7 positives [reports/P_SIGN/run0/p_sign_summary.json]. Thus gain increases can worsen the original segment, but the response is not specific enough to diagnose infeasible references. The G1 gate adds a fourth cost [sealed ✗, kept]: no physics randomisation changes the outcome, so DR budget spent on these segments buys nothing. Practical consequence: the screen is a 1-CPU-second pre-deployment safety filter. There is no validated runtime complement; using gain response as one would contradict the sealed P-SIGN fail.