{"id":"3e48591a-7b0b-4b58-868b-e8440e5638d3","arxiv_id":"2509.09496","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A momentum-based loss term improves physical plausibility of reconstructed and generated human motion without sacrificing accuracy.","lead":"This paper introduces a training loss that makes 3D human motion reconstruction more physically plausible by aligning whole-body linear and angular momentum with real motion data. It reduces foot sliding and jitter in existing models, which matters for animation, VR avatars, and robotics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4c spectrum loss is mathematically redundant with the time-domain angular-momentum loss under DFT/DCT, so the frequency-domain justification in Sec. 3.2 is not actually enforced.","rationale":"The reader's weakest assumption concerns the premise that external torques/forces have small high-frequency content, citing [4]. My concern is adjacent but more concrete: even if that premise holds, Eq. 4c does not implement a high-frequency penalty. Because DFT/DCT are orthogonal transforms, L_S is proportional to the time-domain squared error already present in L_AMo. The ablation's near-identical numbers for L_AMo-only and L_S-only support this reading. This does not overturn the central empirical claim--momentum-based training improves plausibility across tasks and baselines--but it does require a correction to the paper's description of L_S and its physical justification. The reader's verdict of CONDITIONAL remains appropriate; no verdict change is needed, but the condition should include clarifying or revising the spectrum-loss component.","tokens_in":15306,"tokens_out":6550,"duration_ms":83285,"concrete_test":"Recompute L_S using only the high-frequency band (e.g., zero out all DFT/DCT coefficients below a cutoff k0, as in Eq. 10) and retrain the Table 4 GLAMR+LTMo ablation on the same data. If the high-frequency-only L_S reproduces the reported gains, the frequency argument is real; if it behaves like a scaled L_AMo (i.e., Parseval holds and results match Table 4), then Eq. 4c is not an independent spectrum constraint. Additionally, report the exact normalization of F used and verify whether ||F(a)-F(b)||^2 equals a constant times ||a-b||^2 for the actual implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's frequency-domain argument is not operationalized by the proposed L_S. Eq. 4c defines L_S = ||F(AMo(hat R,hat theta)) - F(AMo(R,theta))||^2, where F is stated to be the DFT or DCT. For either of these orthogonal transforms, Parseval's theorem gives L_S = c * ||AMo(hat R,hat theta) - AMo(R,theta)||^2 for a constant c>0. This is exactly the first term of L_AMo in Eq. 4a (up to scaling). Thus L_S is not a frequency-selective constraint; it contains no high-frequency masking or weighting, despite Sec. 3.2 arguing that high-frequency momentum components should be small. The ablation in Table 4 is consistent with this redundancy: L_AMo-only and L_S-only produce nearly identical jitter (15.52 vs 15.53) and FS (5.11 vs 5.19). The central empirical claim that momentum-based training improves plausibility is not invalidated, but the paper's physical/frequency story for L_S is unsupported as written, and the 'spectrum loss' is at best a reweighting of the time-domain angular-momentum term.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TMo, an auxiliary training loss that encourages consistency between predicted/generated human motion and ground-truth whole-body linear momentum (LMo) and angular momentum (AMo), plus a spectrum-based term (L_S) on AMo. The loss is integrated into existing motion models (GLAMR, PhysPT, WHAM, TEMOS) and evaluated on global-trajectory prediction, global motion recovery, and text-to-motion generation. The authors report reduced foot sliding, lower jitter, and improved balance, with comparable accuracy, and provide ablations, weight-sensitivity analysis, a perceptual study, and a stated public code/data release.","tokens_in":15646,"tokens_out":6444,"duration_ms":70695,"significance":"The core idea is appealing and useful: whole-body linear and angular momentum are aggregate physical quantities that couple local joint motion to global root translation/rotation, and the proposed loss is simple and model-agnostic. The multi-task evaluation, component ablations, and perceptual study are strengths, and the paper ships code and data. If the empirical claim holds, this is a low-cost way to improve the plausibility of kinematic motion models. However, the spectrum-loss component is not actually frequency-selective, and the physical/frequency narrative needs substantial correction before the paper can be accepted as written.","major_comments":[{"comment":"The spectrum loss L_S is mathematically redundant with the first term of L_AMo in Eq. (4a). When F is the DFT or an orthonormal DCT, Parseval's theorem gives ||F(AMo(hat R,hat theta)) - F(AMo(R,theta))||^2 = c ||AMo(hat R,hat theta) - AMo(R,theta)||^2 for a positive constant c. Thus L_S contains no frequency weighting or masking and is exactly a scaled version of the ||Delta AMo||^2 component already in L_AMo. The argument in Sec. 3.2 about high-frequency attenuation therefore does not justify L_S as implemented. Table 4 is consistent: the L_AMo-only and L_S-only rows give nearly identical jitter (15.52 vs 15.53) and FS (5.11 vs 5.19). This does not invalidate the momentum-alignment approach, but it removes L_S as a distinct contribution and forces a reinterpretation of the ablations. The authors should either remove L_S, implement a genuinely frequency-selective penalty (e.g., weighting","section":"Sec. 3.3, Eq. (4c)"},{"comment":"The frequency-domain derivation relies on the external claim, citing [4], that external forces and torques have small high-frequency content. The supplementary (D.1) verifies that high-frequency momentum components are small on AMASS, but this is not verified for the evaluation datasets (EMDB, Kungfu, RICH) that contain dynamic or high-impact motions. Because L_S as written is not frequency-selective (see previous comment), the AMASS validation cannot rescue the physical grounding of L_S. Please either verify the assumption on the evaluation data or revise the motivation to describe the loss as aligning full momentum profiles rather than specifically suppressing high-frequency content.","section":"Sec. 3.2"}],"minor_comments":[{"comment":"The abstract claims the loss 'preserves the accuracy' of the recovered motion, but WHAM+LTMo shows slightly worse RTE on both EMDB (4.3 vs 4.1) and RICH (4.4 vs 4.1). The degradation is small, but the claim should be qualified, and ideally the main metrics should be accompanied by error bars or significance tests.","section":"Table 2"},{"comment":"The text says the gaps are 'non-increasing', but the reported numbers are 3.06, 4.12, 2.74, 3.13, which are not monotonically non-increasing. Please correct the description.","section":"Supplementary D.2"},{"comment":"The text says 'We use the discrete Fourier transform F and the discrete cosine transform' for the spectrum loss, but it is unclear whether both are used and how they are combined (e.g., averaged, summed, or used separately). Please specify the exact implementation.","section":"Eq. (4c) / Sec. 3.3"},{"comment":"The PhysPT† baseline is described as using 'the same Transformer architecture as PhysPT, using only position based loss and our global trajectory predictor.' Please clarify the training protocol and how it differs from the original PhysPT, since it is a key comparison.","section":"Sec. 4.1"},{"comment":"There are minor typos, e.g., reference [43] contains 'V ol.3' and the running header 'NGUYEN ET AL: BODY MOMENTUM IN HUMAN MOTION' is repeated. Also, the composite measure m_AB in Sec. 4.4 would benefit from a clearer explanation of the reference direction (baseline at full size).","section":"References / General"}],"recommendation":"major_revision","confidential_remarks":"The redundancy of L_S is a genuine flaw in the paper's technical narrative, but it does not, in my view, undermine the central empirical claim that adding momentum alignment improves plausibility. The paper is likely salvageable after a major revision that reframes or replaces the spectrum term and re-runs the corresponding ablations. I would encourage the authors to release the promised code and to add statistical significance assessments for the main quantitative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll cut to the chase. The momentum-consistency loss (linear and angular) is a simple plug-in that, on the evidence here, actually reduces foot sliding and jitter across GLAMR, WHAM, and TEMOS. The ablations are more thorough than usual—they compare against swing smoothing, point-mass momentum, joint velocity, and direct plausibility objectives. If the numbers hold, this is a useful addition to the motion-recovery toolbox.\n\nThe weak point is the frequency term. Section 3.2 argues that high-frequency momentum components should be small, so the model should be constrained in the frequency domain. But L_S in Eq. 4c is the squared L2 difference between the DFT/DCT of predicted and ground-truth angular momentum. For an orthogonal transform, Parseval's theorem makes this exactly equal to a scaled version of the time-domain L2 difference—i.e., the first term of L_AMo in Eq. 4a. There is no high-frequency masking or weighting. So L_S is not enforcing the paper's stated spectral prior; it is just duplicating part of L_AMo. The ablation confirms it: L_AMo-only and L_S-only give nearly identical jitter (15.52 vs 15.53) and foot sliding (5.11 vs 5.19). This should be fixed—either use a proper frequency-weighted loss or drop L_S and be upfront that momentum matching is the whole story.\n\nOther concerns are smaller. The main quantitative tables (Tables 1–4, 7–10) lack error bars, so significance is unknown; only the generation experiment reports confidence intervals. The perceptual study selects the 40 sequences with the largest baseline-vs-ours gap, so the 70% preference is not a random sample and overstates average improvement. The claim that external forces have small high-frequency content leans on a single biomechanics citation [4] without verification on the diverse AMASS/EMDB/3DPW data. The authors also honestly note the retraining requirement as a limitation. None of these gut the central claim—this is standard supervised regression to ground-truth momentum, not circular—but they do bound how strongly we can trust the details.\n\nBottom line: a practical paper with a genuine contribution and one overstated component. A serious referee should engage with L_S, ask for error bars, and then likely accept after revision. I'd cite it for the momentum loss itself, and would consider bringing it to a reading group focused on human motion as a caution about frequency-domain reasoning.","headline":"Useful momentum loss for motion plausibility, but the spectrum term is Parseval-redundant with the time-domain angular-momentum term, undermining the frequency story.","tokens_in":16074,"tokens_out":4929,"would_cite":true,"duration_ms":53202,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a momentum-matching loss to human motion models reduces foot sliding and jitter while preserving accuracy.","keywords":["human motion","momentum","motion plausibility","loss function","global trajectory","motion recovery","foot sliding","angular momentum"],"falsifier":"Measure the high-frequency content of linear and angular momentum in ground-truth motions that involve hard impacts or rapid direction changes (e.g., parkour landings, quick punches). If these motions exhibit large high-frequency momentum components comparable to the artifacts the loss is meant to suppress, then the premise underlying the spectrum loss is contradicted, and one would predict that L_S either does not help or actively harms plausibility on such motions.","tokens_in":15244,"feed_emoji":"🏃","tokens_out":1902,"duration_ms":26278,"temperature":0.7,"pith_summary":"The paper argues that the local joint rotations and the global root trajectory of human motion are physically coupled, and this coupling is captured by whole-body linear and angular momentum. It introduces a loss term, TMo, that pushes a model's generated momentum profiles to match ground-truth profiles. Adding TMo to existing motion recovery and generation models consistently reduces foot sliding, jitter, and balance errors without hurting accuracy. The loss is simple, requires no architectural changes, and works across multiple baselines and tasks.","feed_headline":"Momentum loss cuts foot sliding and jitter","feed_subtitle":"Matching whole-body linear and angular momentum to real data makes reconstructed human motion more natural without losing accuracy.","key_machinery":"The central object is the whole-body linear momentum and angular momentum of the human body, computed by partitioning the SMPL model into 20 body parts with per-part masses, centroids, and inertia tensors. The loss TMo combines three terms: an angular-momentum matching term (with time derivative), a linear-momentum matching term (with time derivative), and a spectrum term that matches the discrete Fourier/cosine transform of angular momentum to ground truth. This machinery links local joint behavior to global movement without needing explicit force or torque estimation.","core_discovery":"The central claim is that enforcing consistency between generated and ground-truth whole-body linear and angular momentum—computed in a world frame—improves the physical plausibility of reconstructed and generated human motion. The momentum terms aggregate the effect of all joint-level dynamics, so matching them provides a physically grounded bridge between local pose and global displacement. The proposed loss has three parts: matching linear momentum, matching angular momentum, and matching the frequency spectrum of angular momentum to suppress unnatural high-frequency content. Experiments on global trajectory recovery, full motion recovery, and text-to-motion generation show that the loss","pith_inferences":["A natural extension the paper leaves implicit is using the momentum-matching loss as a self-supervised test-time refinement objective, since momentum can be computed from the model's own outputs without ground-truth labels.","The momentum plausibility detector (based on high-frequency components) could be repurposed as a standalone evaluation metric or a filtering step for motion datasets, not just a training loss.","The same formulation could transfer to other articulated body models or even non-human characters, as long as a part-based mass and inertia model can be defined.","The loss might also serve as a regularizer in motion prediction tasks, where the future trajectory must remain dynamically consistent with the evolving pose."],"forward_implications":["If the central claim holds, any kinematic motion model—reconstruction, prediction, or generation—can be made more physically plausible by adding the TMo loss during training, without redesigning the architecture.","The loss yields consistent improvements across diverse baselines (GLAMR, WHAM, PhysPT, TEMOS) and across datasets including in-the-wild and acrobatic motions, suggesting the coupling is general rather than task-specific.","The method performs better in low-data regimes, implying the momentum constraint acts as a useful inductive bias that reduces the amount of motion data needed.","The frequency-spectrum component provides a new, physically motivated detector of implausible motion: sequences with large high-frequency momentum components are likely unrealistic."],"fun_headline_variants":["Momentum constraint cuts foot slide and jitter in motion","Whole-body momentum loss improves motion plausibility","Momentum matching links local pose to global movement","Momentum loss reduces foot sliding and jitter, improves balance","Momentum tie-in trims foot slide and jitter"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The frequency-domain justification for the spectrum loss assumes that external forces and torques acting on the body have small high-frequency content, so if a motion involves sharp impacts or very fast force changes (e.g., acrobatic landings), that specific loss term's physical grounding weakens.","fun_headline_variants_meta":{"raw":{"variants":["Momentum constraint cuts foot slide and jitter in motion","Whole-body momentum loss improves motion plausibility","Momentum matching links local pose to global movement","Momentum loss reduces foot sliding and jitter, improves balance","Momentum tie-in trims foot slide and jitter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3368,"prompt_tokens":698,"completion_tokens":2670,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":2589}},"tokens_in":442,"tokens_out":2670,"duration_ms":21966,"temperature":1.0,"reasoning_tokens":2589,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:59:25.239786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the high-frequency content of linear and angular momentum in ground-truth motions that involve hard impacts or rapid direction changes (e.g., parkour landings, quick punches). If these motions exhibit large high-frequency momentum components comparable to the artifacts the loss is meant to suppress, then the premise underlying the spectrum loss is contradicted, and one would predict that L_S either does not help or actively harms plausibility on such motions.","supporting_citations":[],"review_version":1}