cs.CVSep 29, 2026

Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of Generative Motion Models

Authors: Zhiyuan Li, Wenyan Yang, Pekka Marttinen, Joni Pajarinen

Organizations: Department of Electrical Engineering and Automation, Aalto University, Finland · Department of Computer Science, Aalto University, Finland

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Cross-skeleton motion generation trains generative models to carry action structure and motion intention from one body to another. Yet a target motion that shows the right action has two explanations that the training data cannot tell apart: the model transferred the source clip, or it recovered a typical motion for the requested action. We show that this ambiguity is structural rather than incidental: under standard generative objectives, the source-conditioned retargeting map is non-identifiable in sparse heterogeneous motion domains. Unpaired distribution matching yields gauge non-identifiability: the latent spaces of different skeletons can be transformed relative to one another without changing the training evidence, so different source-conditioned maps fit it equally well. Sparse paired supervision admits the complementary failure mode, \emph{conditional-mean degeneration}: when clips are paired only by action, squared-error training converges to an average target motion that ignores the source clip. To make the missing evidence observable, we introduce Source-Instance Fidelity (SIF), a diagnostic that tests whether outputs differ from one another the way their source clips do, with the target skeleton and action held fixed. Under this diagnostic, methods that succeed at the standard action-level test on animal motion data often sit at the source-blind floor, while the methods that rise above it retain only a partial relational signal. Retargeting therefore needs objectives and evaluations that can identify the source-conditioned map it claims to learn. Project page: https://cross-skeleton-retargeting.netlify.app/.

Figures & tables

Figure 1: Action-level success does not identify the source-conditioned map. Two source clips with the same action can lead either to distinct target motions under a source-preserving map or to a shared target-action habit under source-blind recovery. Standard action-level AUC accepts both explanations; Source-Instance Fidelity tests whether the source-instance relation survives the change of body.
Figure 2: Cell-paired squared-error regression collapses toward the target-cell prototype. Panel A illustrates random target-cell pairing: several admissible target clips define a cell, but the squared-error optimum is the cell mean prototype rather than a source-conditioned motion. Panel B reports the controlled ladder with known source-conditioned transport T⋆T^{\star} ; from M=2M=2 to M=50M=50 pairs per cell, predicted output variance stays at about 2% of the oracle within-cell variance Var(T⋆(xa))\mathrm{Var}(T^{\star}(x_{a})) . The shaded band is the three-seed standard deviation.
Figure 3: Source-Instance Fidelity for the fourteen evaluated methods on the 1,891 Truebones triples. Filled circles use raw scoring and open squares length-controlled scoring; intervals are 95% bootstrap intervals that resample whole source skeletons, and nn is the number of triples a method supports. The shaded band marks ∣ρ\textscsif∣<0.10\left|\rho_{\textsc{sif}}\right|<0.10 . Eight methods do not stand out from shuffled pairings under either scoring, and Motion2Motion-BVH stays near zero.
Figure 4: Qualitative non-identifiability panorama. A single Bird →\to KingCobra attack query is shown across evaluated methods. All rows receive the same source clip, target skeleton, and action request, yet the generated target motions differ visibly across methods. The figure is qualitative evidence for the ambiguity isolated by SIF: action-consistent target motion can be produced without identifying a unique source-conditioned map.
MethodOverall n/600n/600Held-out n/200n/200Overall AUCHeld-out AUC
ANCHOR6002000.7570.681
random-same-cluster5701770.769 (+0.012)0.678 (-0.003)
random-same-exact-action221570.883 (+0.126)0.923 (+0.242)
Parenthetical Δ\Delta values are method AUC minus ANCHOR. random-same-cluster matches ANCHOR within ±0.012\pm 0.012 on both splits. random-same-exact-action exceeds ANCHOR by +0.126+0.126 overall and +0.242+0.242 held-out, but only on the 221/600221/600 and 57/20057/200 coverage where exact-action positives exist; the coverage qualifier travels with the gap.
Table 1: Label-only retrieval can match or exceed ANCHOR’s action-level AUC without solving the retargeting problem. AUC is measured under the same cluster-tier Procrustes protocol used for ANCHOR. The exact-action reference has restricted coverage, so its larger gaps are only defined on the rows where the target library contains an exact-action candidate.
Figure 5: Action-level AUC cannot distinguish source-conditioned transport from label-based retrieval. ANCHOR retrieves from the target library using action information and motion descriptors, while the two random references use only label constraints. Their AUC values remain in the same action-level regime, showing that the standard score can accept target-action recovery even when source-instance transfer has not been established. The exact-action reference is coverage-limited to 221/600221/600 overall and 57/20057/200 held-out queries.
 SIFAction AUCVariationRealistic
True retargeted counterpart+0.960+0.9600.9800.5297.2%
Random same-action clip−0.008-0.0080.9780.4197.5%
Unpaired objective (Theorem 1 class)+0.203+0.2030.4990.001100.0%
Averaging objective (Proposition 1.1 class)+0.379+0.3790.8200.0553.0%
Model trained on true pairs+0.900+0.9000.9670.5294.6%
Table 2: Human-to-G1 evaluation on identical test items (90 action groups, 355 clips). SIF uses raw scoring; variation is the median ratio of output spread to source spread; realistic is the share of outputs whose nearest real G1 pose lies within the 95th percentile of held-out real motion.
Appendix figures & tables47 assets

Supplementary material from the paper’s appendix.

Appendix

QuantityValue
Skeletons70
Total clips616
Coarse clusters10
Exact actions90
Skeleton-action cells6,300 ( 70×9070\times 90 )
Occupied cells413 (6.6%)
Appendix
Table 3: Truebones zoo regime. SIF is evaluated on source groups with at least three clips; the main evaluation pairs each group with every target skeleton that performs the same action, and the original evaluation pairs it with a single target.
MethodFamilyTrain skeletons
AnyTopsource-conditioned diffusion70
ACE-Tmotion-space adversarial70
ACE-Imotion-space adversarial60
MoReFlow-Tflow matching70
MoReFlow-Iflow matching60
AL-Flowlabel-conditional flow, no source motion60
Appendix
Table 4: Method roster. Train-skeleton counts distinguish transductive evaluation on all skeletons from inductive evaluation with held-out target skeletons.
MethodRelation to the formal results
AnyTopclosest to Theorem 1 ; adds rank and view-consistency terms
AL-Flow, AL-Flow-Src, AL-Flow-Src-Gclosest to Proposition 1.1 ; trained with flow matching rather than squared error
MoReFlow-T, MoReFlow-Ipartly within Theorem 1 ; adds pairing by matched motion descriptors and a descriptor loss
DPG-SB-v3partly within Theorem 1 ; adds adversarial and cycle terms in the latent space
ACE-T, ACE-Ioutside both classes; adversarial and source-feature losses
ANCHOR, random-same-cluster, random-same-exact-actionretrieval rules with no trained objective (Corollary 1.1 )
Appendix
Table 5: Relation of each evaluated method to the formal results.
Figure 6: Decoder non-degeneracy check. Rotation-induced changes are compared with matched-magnitude noise in the latent space; ratios above one indicate that the decoder distinguishes at least some directions on the gauge orbit.
MethodSeedChange under rotationChange under noiseRatioVerdict
ACE-T (3 seeds)420.2400.2131.93non-degenerate
 430.2190.1831.78non-degenerate
 440.2090.1712.04non-degenerate
ACE-I (3 seeds)420.1210.0892.35non-degenerate
 430.1280.0711.98non-degenerate
 440.1410.0871.92non-degenerate
Appendix
Table 6: Rotation versus noise perturbations on the 130 queries of the 37-triple intersection. The ACE ablation with the motion-space adversarial loss removed is the boundary case near the noise floor, while the other evaluated methods make the orbit visible after decoding.
GroupSeed pairShare explained by rotationCosine after rotation
ACE-T42 vs 430.5220.893
ACE-T42 vs 440.7030.900
ACE-T43 vs 440.5850.905
ACE-I42 vs 430.6890.935
ACE-I42 vs 440.7120.928
ACE-I43 vs 440.7750.941
Appendix
Table 7: Orthogonal Procrustes alignment across seeds. The table reports how much latent variation can be explained by an O(d)O(d) alignment across the SIF-intersection queries.
Figure 7: Cross-seed Procrustes alignment. High explained fractions indicate that latent spaces differ partly by approximately orthogonal transformations, matching the relative-gauge mechanism in Theorem 1 .
MethodSample shapeSpectral flatnessEffective rank // 256
ACE without adversarial loss (seed 42)3,376 ×\times 2560.00011.29
ACE-T (seed 42)3,376 ×\times 2560.00012.13
ACE-I (seed 42)3,376 ×\times 2560.00193.00
AL-Flow1,040 ×\times 2560.00666.80
MoReFlow-I3,823 ×\times 2560.059312.84
MoReFlow-T3,823 ×\times 2560.063114.99
Appendix
Table 8: Effective-rank diagnostic on latent distributions. Higher effective rank indicates richer latent variation, but rank alone does not imply that the source-conditioned map has been identified.
Figure 8: Effective rank across latent representations. The diagnostic separates representation richness from source-instance fidelity.
Figure 9: Decoder non-degeneracy and SIF are distinct. A method may expose latent perturbations after decoding while still lying near the source-blind floor under SIF.
MethodnnRaw SIF [95% CI]ppLength-controlled SIF [95% CI]ppVariationReading
ACE-I1,891++ 0.420 [ ++ 0.31, ++ 0.53]<< 0.001++ 0.248 [ ++ 0.13, ++ 0.38]<< 0.0010.03above floor
MoReFlow-T1,891++ 0.255 [ ++ 0.19, ++ 0.33]<< 0.001++ 0.246 [ ++ 0.17, ++ 0.32]<< 0.0010.20above floor
MoReFlow-I1,891++ 0.257 [ ++ 0.19, ++ 0.33]<< 0.001++ 0.244 [ ++ 0.18, ++ 0.32]<< 0.0010.16above floor
ACE-T1,891++ 0.336 [ ++ 0.24, ++ 0.44]<< 0.001++ 0.180 [ ++ 0.05, ++ 0.31]0.0090.02above floor
Motion2Motion-Direct1,891++ 0.143 [ ++ 0.08, ++ 0.22]<< 0.001++ 0.129 [ ++ 0.06, ++ 0.20]<< 0.0011.52above floor
Motion2Motion-BVH1,872++ 0.032 [0.000, ++ 0.07]0.024++ 0.032 [ −- 0.003, ++ 0.07]0.0160.79near floor
Appendix
Table 9: SIF on the main evaluation (1,891 triples). Raw and length-controlled SIF with 95% intervals that resample whole source skeletons; pp is the one-sided shuffle-test pp -value; variation is the median ratio of output spread to source spread under raw scoring; nn is the number of triples a method supports. The last row is the ACE ablation of Appendix I , not an evaluated method.
MethodnnRaw SIFppLength-controlled SIFppVariation
MoReFlow-T49++ 0.2030.011++ 0.2200.0050.25
ACE-T49++ 0.385<< 0.001++ 0.2050.0130.02
ACE-I49++ 0.484<< 0.001++ 0.1780.0220.02
MoReFlow-I49++ 0.1370.061++ 0.1600.0380.28
Motion2Motion-Direct49++ 0.0530.290++ 0.0910.1461.69
ANCHOR49++ 0.0140.380++ 0.0350.2260.00
Appendix
Table 10: SIF on the original evaluation (49 triples, one target per source group). Columns as in Table 9 ; here no source clip is shared between triples.
Methodeach triple aloneby source skeletonby target skeletonby action
ACE-I0.526 [0.351, 0.681]0.526 [0.323, 0.709]0.526 [0.365, 0.671]0.526 [0.242, 0.629]
ACE-T0.453 [0.293, 0.602]0.453 [0.284, 0.618]0.453 [0.274, 0.605]0.453 [-0.009, 0.619]
MoReFlow-T0.192 [-0.003, 0.380]0.192 [-0.023, 0.386]0.192 [-0.031, 0.405]0.192 [-0.039, 0.357]
random-same-exact-action0.185 [0.046, 0.326]0.185 [0.069, 0.304]0.185 [0.027, 0.354]0.185 [0.000, 0.272]
Motion2Motion-Direct0.131 [-0.061, 0.317]0.131 [-0.098, 0.385]0.131 [-0.075, 0.307]0.131 [-0.720, 0.265]
random-same-cluster0.104 [-0.095, 0.298]0.104 [-0.057, 0.279]0.104 [-0.115, 0.315]0.104 [-0.272, 0.189]
Appendix
Table 11: Clustered bootstrap intervals for SIF. Clustering by source skeleton, target skeleton, or action tests whether the positive conclusions are driven by a narrow subset of triples.
Figure 10: Clustered confidence intervals for SIF. The figure visualizes the dependence-aware intervals reported in Table 11 .
EvaluationScoringACE-T minus MoReFlow-TACE-I minus MoReFlow-I
1,891 triplesraw++ 0.081 [ −- 0.02, ++ 0.19], pp =0.139++ 0.163 [ ++ 0.05, ++ 0.28], pp =0.008
1,891 tripleslength-controlled−- 0.066 [ −- 0.19, ++ 0.06], pp =0.325++ 0.004 [ −- 0.10, ++ 0.11], pp =0.935
49 triplesraw++ 0.182 [ −- 0.06, ++ 0.44], pp =0.156++ 0.347 [ ++ 0.11, ++ 0.60], pp =0.009
49 tripleslength-controlled−- 0.015 [ −- 0.21, ++ 0.19], pp =0.888++ 0.018 [ −- 0.21, ++ 0.24], pp =0.880
Appendix
Table 12: Paired SIF differences on identical triples, with 95% intervals that resample whole source skeletons and two-sided sign-flip pp -values that flip all triples of a source skeleton together.
 One tripleAverage over kk triples
Method(three clips)k=25k=25k=100k=100k=500k=500k=1,000k=1{,}000
AnyTop0.600.1220.0600.0210.012
ACE-T0.590.1240.0600.0220.013
ACE-I0.560.1100.0540.0190.012
MoReFlow-T0.570.1230.0590.0230.013
random-same-exact-action0.310.0890.0460.0170.010
Appendix
Table 13: Noise of a single triple and stability of the average. The first column is the standard deviation of SIF across three-clip subsets of five-clip triples; the other columns are standard deviations of the average over kk randomly drawn triples.
 Independent noise, three seeds  
Methodseed 1seed 2seed 3Mean ±\pm sdShared noise
AnyTop−- 0.033++ 0.040−- 0.098−- 0.031 ±\pm 0.069++ 0.243
AL-Flow−- 0.066−- 0.043++ 0.066−- 0.015 ±\pm 0.070++ 0.002
AL-Flow-Src++ 0.070++ 0.049−- 0.140−- 0.007 ±\pm 0.115++ 0.199
AL-Flow-Src-G−- 0.028++ 0.048++ 0.075++ 0.031 ±\pm 0.053++ 0.239
DPG-SB-v3−- 0.058++ 0.138++ 0.052++ 0.044 ±\pm 0.098++ 0.064
Appendix
Table 14: Generation noise control on the original 49 triples. Independent noise uses three seeds; shared noise reuses one noise sample within each triple.
MethodMin pairsMedian pairsMax pairsnn triples
ACE-I33549
ACE-T33549
AL-Flow33545
AL-Flow-Src33538
AL-Flow-Src-G33549
ANCHOR33549
Appendix
Table 15: Pair counts per SIF triple. The minimum and median values are fixed by the eligibility rule; the few larger cells determine how much within-cell geometry is available.
Figure 11: Pair-count distribution for SIF-eligible triples.
Figure 12: SIF versus output variation (the diversity ratio RR ). High output variation is not sufficient for source-instance preservation, because the geometry must align with the source-side geometry after the target skeleton and action are fixed.
MethodSIF ρ\rhoL-SIF ρ\rho
ACE trained with zsrc=0z_{\text{src}}=00.4230.037
ACE-I (3 seeds avg)0.4100.049
ACE-T (3 seeds avg)0.4080.088
ACE without adversarial loss (3 seeds avg)0.2140.119
MoReFlow-T0.192-0.200
MoReFlow-I0.076-0.205
Appendix
Table 16: SIF and L-SIF for each method. L-SIF is a latent-space diagnostic and is therefore reported only where the representation is comparable across queries.
Figure 13: SIF and L-SIF decoupling. Source-instance fidelity in decoded motion need not coincide with a positive latent-space correlation.
Figure 14: Synthetic 2×22\times 2 SIF calibration. The oracle uses the known source-conditioned transport T⋆(xa)T^{\star}(x_{a}) ; the random-clip and Gaussian references deliberately ignore the source instance.
OutputSIF (95% CI)ppVariation
Oracle (output =T∗(xa)=T^{*}(x_{a}) )++ 0.993 [ ++ 0.992, ++ 0.995]<< 0.0010.94
Random clip from cell++ 0.002 [ −- 0.027, ++ 0.029]0.4560.96
Random Gaussian noise++ 0.021 [ −- 0.008, ++ 0.054]0.1641.3×1051.3{\times}10^{5}
Synthetic 2 ×\times 2 paired-dense setting, 168 groups of three to six source clips. Another 168 groups are left out because their source clips are identical once rotation and scale are removed, so SIF is undefined there. The oracle saturates SIF; both source-blind references sit at zero. Noise has no motion structure, so its variation is very large.
Appendix
Table 17: Synthetic SIF calibration. The oracle lies near perfect source-instance preservation, while source-blind references remain near zero.
MM pairs / cellσpred2/σoracle2\sigma^{2}_{\text{pred}}/\sigma^{2}_{\text{oracle}}MSE /σoracle2/\sigma^{2}_{\text{oracle}}nseedsn_{\text{seeds}}
20.020 ±\pm 0.0061.45 ±\pm 0.083
40.014 ±\pm 0.0021.29 ±\pm 0.033
80.020 ±\pm 0.0061.34 ±\pm 0.083
160.015 ±\pm 0.0021.31 ±\pm 0.123
320.017 ±\pm 0.0011.22 ±\pm 0.053
500.016 ±\pm 0.0031.26 ±\pm 0.093
Appendix
Table 18: Conditional-mean degeneration in the controlled 2×22\times 2 setting. The predicted variance remains a small fraction of oracle variance even when the number of random target-cell pairs increases.
DomainCluster classifier accuracynn correct // total
In-domain (train skeletons)99.8%543 // 544
Held-out (reserved cross-skeleton)61.5%59 // 96
Replacing the ground-truth cluster label with the predicted cluster pulls Procrustes performance below random on exact-tier and on a hand-curated query subset. The in-domain row is the accuracy on the clips the classifier was trained on.
Appendix
Table 19: ANCHOR action-cluster predictor. The held-out row shows that the cluster prediction itself is imperfect, which is why ANCHOR is interpreted as a comparator rather than an oracle.
Figure 15: Per-cluster ANCHOR matches. The comparator’s behavior varies across action clusters, reinforcing that it characterizes a floor rather than a universal retargeting rule.
Subsetnn triples (cluster-eligible)Cluster-tier AUC (95% CI)
All30,0290.824 [0.820, 0.829]
In-distribution (same train pair)21,8380.845 [0.841, 0.849]
Mixed (one held-out skeleton)7,6220.774 [0.762, 0.787]
Held-out (both cross-skeleton)5690.711 [0.669, 0.758]
ANCHOR on the 30,497-pair full enumeration. Cluster-eligibility filter removes triples whose target skeleton has no in-cluster positives. The mixed subset is the held-out case.
Appendix
Table 20: Full ANCHOR enumeration across cluster-eligible triples. The held-out subsets show that strong action-level ranking can arise from target-library label structure without identifying a source-conditioned map.
Figure 16: Enumeration breakdown for ANCHOR. The large enumeration separates in-distribution, mixed, and held-out skeleton regimes.
VariantConditioning channelsCluster AUC
AnyTopself-supervised source-conditioned diffusion0.465
Precursor action-conditioned generatoraction-conditioned generation0.485
Precursor latent-bridge generatorlatent bridge objective0.483
DPG-SB-v3latent bridge objective; no decoded-motion penalties0.447
AL-Flowcluster + exact-action labels0.538
AL-Flow-SrcAL-Flow + source motion + skeleton identity0.542
Appendix
Table 21: Action-level AUC for generated variants and supervision-matched comparators. These values are not interpreted as source-preserving transfer; they show how much action-level evidence can be explained without SIF.
Cluster-tier held-out AUC (average of folds 42 and 43)
MethodProcrustesZ-DTWQ-comp
Action oracle0.8730.8460.894
Self-positive reference0.8600.8470.909
ANCHOR0.6810.7040.728
Cluster-Classifier Retrieval0.6460.6690.693
Q-Retrieval (Q-feature only)0.4530.5230.597
Appendix
Table 22: Master action-level AUC comparison on held-out queries. The table reports cluster-tier AUC under three distances: Procrustes trajectory distance; Z-DTW, dynamic time warping on z-normalized body-part trajectories; and Q-comp, a combination of centre-of-mass path, foot-contact timing, cadence, and limb-usage differences. Library-based references are listed separately from generative methods.
Figure 17: Action-level AUC band. Many methods occupy a narrow action-ranking band, which motivates SIF as the missing source-instance axis.
  ACE-TModel trained with zsrc=0z_{\text{src}}=0
Source at generationOutput lengthSIFppVariationSIFppVariation
realcopied++ 0.385<< 0.0010.020++ 0.419<< 0.0010.023
realfixed++ 0.294<< 0.0010.003++ 0.2130.0060.004
removedcopied++ 0.300<< 0.0010.009++ 0.299<< 0.0010.007
removedfixed0.0001.0000.0000.0001.0000.000
shuffledcopied−- 0.1920.9870.020−- 0.1650.9700.023
Appendix
Table 23: Removing or shuffling ACE’s source input at generation (49 original triples). Removed means zsrcz_{\text{src}} set to zero at generation; shuffled means each output receives the source of another clip in the same triple; fixed means every output is generated with 64 frames. pp is the shuffle-test pp -value; variation as in Table 9 .
  With adversarial lossWithout  
ScoringTriplesSIFVariationSIFVariationDifference [95% CI]pp
raw, length copied49++ 0.4070.017–0.022++ 0.2280.271–0.906++ 0.179 [ ++ 0.04, ++ 0.33]0.022
length fixed at generation49++ 0.2920.002–0.004++ 0.2660.073–0.107++ 0.026 [ −- 0.09, ++ 0.15]0.677
length fixed at generation1,891++ 0.3290.002–0.003++ 0.3290.078–0.1110.000 [ −- 0.05, ++ 0.06]0.999
Appendix
Table 24: ACE-I with and without the adversarial loss, trained identically otherwise (three seeds each). SIF is averaged over seeds; variation gives the range of the per-seed medians; the difference is paired over triples, with a 95% interval that resamples whole source skeletons.
 Actual outputsIdentical copies,
MethodRawStretched to 64 framesstretched to 64 frames
ACE-T0.0200.2310.149
ACE-I0.0200.1350.142
MoReFlow-T0.2470.2170.093
MoReFlow-I0.2760.2790.078
Appendix
Table 25: Output variation after stretching every output to 64 frames (49 original triples). Identical copies are one output cut to the lengths of its siblings, so any variation they show comes from stretching alone.
MethodQueriesCloser to the target poolMean distance to sourceMean distance to target poolRatio
ACE-I130103 (79.2%)3.4492.2151.56
ACE-T13099 (76.2%)3.6972.6501.39
Distance computed in a four-dimensional kinematic feature space (centre-of-mass displacement, centre-of-mass variance, mean velocity, foot-contact density). 76–79% of ACE outputs are closer to a target-pool training clip than to the corresponding source clip, weakening source-memorisation as an explanation of ACE’s SIF lift.
Appendix
Table 26: ACE nearest-neighbor leakage check. Distances are computed in a kinematic feature space; the majority of ACE outputs are closer to the target pool than to the source clip.
Figure 18: Human-to-robot renders. (A) One LAFAN1 dance routine (10 s) retargeted by GMR to six humanoid robots; five key poses per strip, earliest faded. Each strip is framed to its robot, so compare poses, not sizes; the base heights give the true scale. (B) Two performers dancing the same routine, and what each model produces for each of them on the Unitree G1. The true retargets and the model trained on true pairs differ between the performers. The averaging and unpaired objectives produce nearly the same motion for both, and that motion barely changes over time. Model outputs are predicted body positions, drawn as the nearest G1 pose; the fit error is listed under each row. The 5 to 6 cm errors of the two collapsed rows mean their raw predictions are less robot-like than drawn. The two performers are not aligned in time, so part of the difference in the top rows is timing rather than style.
 Raw SIFLength-controlled SIFAction AUCVariationRealistic
True retargeted counterpart++ 0.960++ 0.9590.9800.51897.2%
Random same-action clip−- 0.008−- 0.0680.9780.41497.5%
Unpaired objective++ 0.203++ 0.1540.4990.001100.0%
Averaging objective++ 0.379++ 0.3430.8200.04853.0%
Model trained on true pairs++ 0.900++ 0.6980.9670.52194.6%
Appendix
Table 27: Human-to-G1 study, all four measures on identical test items (90 groups, 355 clips), with SIF under both scorings. Variation is the median ratio of output spread to source spread; realistic as in Table 2 .
PredictionThresholdResult
The true retarget is recognizedSIF ≥0.80\geq 0.80 , interval above 019 of 24; every interval above 0; misses 0.768–0.799
A random same-group clip sits at the floor90% interval within ±0.20\pm 0.2012 of 24; 12 of 12 in the denser setting
The averaging objective loses variationvariation Q≤0.35Q\leq 0.35 , upper bound <0.50<0.5024 of 24
The unpaired objective loses variationvariation Q≤0.35Q\leq 0.35 , upper bound <0.50<0.5022 of 24; missed with variation Q=0.41Q=0.41
The true-pair model recovers bothSIF ≥0.60\geq 0.60 and variation Q≥0.70Q\geq 0.7024 of 24; SIF 0.77–0.99
Appendix
Table 28: Six-robot predictions, fixed before any trained model was scored, and their outcomes over 24 evaluations (six robots, two settings, two scorings). QQ is a model’s output variation divided by that of the true counterpart, matched group by group; intervals are group bootstraps.
   True retargetRandom clipUnpairedAveragingTrue-pair model
RobotSettingScoringSIFSIFVariation ( QQ )Variation ( QQ )SIFVariation ( QQ )
Unitree G1three-clipraw++ 0.988++ 0.0660.0590.286++ 0.9871.00
Unitree G1three-cliplength-controlled++ 0.987++ 0.0660.0590.286++ 0.9871.00
Unitree G1denserraw++ 0.966−- 0.0610.0120.035++ 0.9650.99
Unitree G1denserlength-controlled++ 0.966−- 0.0580.0120.035++ 0.9650.99
Booster T1three-clipraw++ 0.799++ 0.0240.0550.347++ 0.8450.99
Appendix
Table 29: Six-robot results for every robot, setting, and scoring. SIF columns give the correlation; QQ columns give output variation relative to the true counterpart.
 Raw SIFAction AUCVariationRealistic
Planned run++ 0.2090.500diverged0.0%
Re-run, seed 1++ 0.6960.8500.56114.4%
Re-run, seed 2++ 0.5280.5510.09378.0%
Re-run, seed 3++ 0.5740.5130.1010.0%
Appendix
Table 30: An adversarial objective in the style of ACE on the human-to-G1 data, without pairs. The planned run diverged; the re-runs use ACE’s optimizer settings and are post hoc.
MethodRaw SIFAction AUCVariationRealistic
ACE-I++ 0.4840.5390.0291.8%
ACE-T++ 0.3850.5380.0292.4%
MoReFlow-T++ 0.2030.5960.2599.4%
MoReFlow-I++ 0.1370.5780.2899.4%
random-same-exact-action++ 0.0960.7990.00100.0%
random-same-cluster++ 0.0730.8130.25100.0%
Appendix
Table 31: Truebones methods on the 49 original triples: SIF (raw), action AUC, output variation, and realism on identical test items. Realism is undefined for Motion2Motion-BVH, whose outputs use a different joint layout.
 Raw SIFAction AUCVariation ( QQ )Realistic
Three clips per group
True retargeted counterpart+0.87 (+0.77–+0.99)0.61 (0.60–0.63)1.00 (1.00–1.00)99 (99–99)%
Random same-group clip+0.08 (+0.02–+0.18)0.62 (0.61–0.63)0.73 (0.69–0.77)99 (99–100)%
Unpaired objective+0.24 (+0.11–+0.36)0.50 (0.49–0.52)0.15 (0.02–0.41)98 (89–100)%
Averaging objective+0.11 (+0.07–+0.20)0.60 (0.58–0.63)0.29 (0.26–0.35)46 (33–83)%
Model trained on true pairs+0.89 (+0.85–+0.99)0.61 (0.60–0.63)1.00 (0.99–1.00)98 (97–99)%
Appendix
Table 32: Six-robot study, four measures on identical test items: mean over the six robots, with the range in parentheses. Variation is QQ , a model’s output variation divided by that of the true counterpart. Action AUC and realism were computed after the pre-registration.
Figure 19: A source-clip substitution view on one Bird →\to KingCobra attack triple. The first three columns, labeled source 1 to source 3, are different Bird attack clips, while the target skeleton and action are fixed; the three columns on the right, labeled output 1 to output 3, show the corresponding KingCobra motions generated by each row. Each row is labeled by its source-instance failure mode: latent collapse for AL-Flow, where source identity is lost before decoding; decoder collapse for AnyTop, where latent variation is not expressed in the output; partial source-conditionality for MoReFlow-T, the clearest output-level source dependence among the displayed rows; and collapsed variation for ACE-I, whose outputs keep a positive correlation with their sources but remain nearly identical.
Figure 20: Full qualitative grid A. Three source clips are compared against target outputs from the evaluated methods on a fixed source-target-action triple.
Figure 21: Full qualitative grid B. The same layout is used to reveal whether output variation follows the source clips after target skeleton and action are fixed.
Figure 22: Full qualitative grid C. The grid complements Figure 19 by showing all evaluated methods on an additional held-out triple.

Explore similar work

CardsList

  1. Retargeting Motions to Diverse Skeletons via Learnable Flattening

    Sep 29, 2026Kia-Jüng Yang, Fabian H. Sinz, Paweł A. PierzchlewiczCross-Embodiment RetargetingSkeleton

    Institute of Computer Science, University of Göttingen · Campus Institute Data Science, University Göttingen · Pantomim P.S.A

  2. Human2Humanoid: Physics-Aware Cross-Morphology Motion Retargeting for Humanoid Robots

    Jun 2, 2026Tianchen Huang, Feiyang Yuan, Junchi Gu +5Cross-Embodiment RetargetingHuman Motion

    Institute of Humanoid Robots, Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China, Hefei, Anhui 230026, China

  3. A Unified Conditional Flow for Motion Generation, Editing, and Intra-Structural Retargeting

    Date pendingJunlin Li, Xinhao Song, Siqi Wang +2Text-To-Motion GenerationCross-Embodiment Retargeting

    ByteDance, San Jose, USA