{"id":"f83a3ead-d0b4-4e6c-b670-be8c2a7c37db","arxiv_id":"2608.03227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A flow-matching pose prior, kept frozen, provides a Jacobian-based directional score that modulates tracking rewards and improves RL-based humanoid motion tracking, especially for dynamic acrobatic skills.","lead":"This paper presents a frozen flow-matching pose prior for humanoid tracking, whose Jacobian defines a Pose Geometry Score (PGS) that rates how well a robot's joint changes align with learned pose geometry. It uses this score to shape the reinforcement learning reward and reports faster, more accurate tracking for dynamic skills like backflips and cartwheels, both in simulation and on a real robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The geometric interpretation of PGS assumes the learned denoiser Jacobian J_phi matches the population-optimal conditional-covariance Jacobian J* at rollout query points; no quantitative check of this faithfulness is provided, so the central mechanism may be an arbitrary smoothness regularizer.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the learned denoiser Jacobian's faithfulness to the population-optimal conditional covariance is asserted but not verified at online query points. I agree that this is the most central threat to the paper's conceptual claim. The empirical improvements could survive even if PGS is an arbitrary smoothness regularizer, but then PFM-HR's novelty—'pose geometry'—would be largely rhetorical. The t_eval selection issue is real and worth fixing, but it primarily affects the general-tracking headline numbers, not the mechanism; the single-trajectory experiments still show gains on dynamic skills. The absence of code/data is a reproducibility concern, not a correctness one. Appendix B's perturbation studies are well designed but only establish internal consistency and sensitivity of the learned score; they do not ground it in the data distribution. The proposed check directly estimates C_t(z) from the pretraining corpus and compares it to J_phi^T J_phi at policy-visited query points, which is the minimal experiment that would settle the assumption. Since the reader already conditioned acceptance on this concern, no verdict change is needed; the paper remains a conditional accept pending this verification.","tokens_in":19160,"tokens_out":6233,"duration_ms":72720,"concrete_test":"Take a held-out set of rollout transitions (q_{k-1}, d_k) from the trained PFM-HR policy on dynamic tasks such as Backflip and Sideflip, not from the prior training corpus. For each transition, fix t_eval=0.75, draw ε, form z = 0.75 q_{k-1} + 0.25 ε, and estimate C_t(z) empirically: sample S≈10^4 clean poses x_s from BONES-SEED, weight them by N(z; 0.75 x_s, 0.25^2 I), and compute the weighted covariance. Then compare the learned geometric tensor G_phi(z) = J_phi(z)^T J_phi(z) with the population target (0.75/0.25^2)^2 C_t(z)^2 using normalized Frobenius error and cosine overlap of top eigenvectors. Also compare PGS rankings of original vs rotated/permuted d_k against rankings produced by the empirical target. If the overlap is below ~0.5 or the Frobenius relative error exceeds ~50%, the geometric interpretation fails; if they match, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's conceptual contribution—that PGS measures alignment with pose-prior co-variation modes—rests entirely on the identity in Eq. (11) and Appendix A: at the population optimum, J*_t(z) = (t/(1-t)^2) Cov[x|z_t=z], so PGS in Eq. (16) equals a squared-covariance-weighted projection of the joint-change direction d_k. But PFM-HR does not use J*; it uses the Jacobian J_phi of a finite-capacity residual MLP. The paper provides no evidence that J_phi approximates J* at the actual online query points (q_{k-1}, t_eval). Appendix B's robustness studies perturb d at fixed queries and measure consistency, discriminability, and magnitude-invariance of the learned score; these are internal consistency checks and would pass even for an arbitrary smooth network whose Jacobian has little to do with the data covariance. The derivation in Appendix A explicitly says the interpretation is approximate 'to the extent that its input Jacobian agrees with the population-optimal Jacobian,' but the extent is never quantified. If J_phi's directional response diverges from the true conditional covariance, PGS is not a geometry-guided signal; it is just another reward-shaping term, and the central claim that the prior reveals pose geometry is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes PFM-HR, a frozen flow-matching pose prior trained on unordered humanoid poses, and uses the Jacobian of its clean-pose denoiser to define a Pose Geometry Score (PGS). PGS is inserted into an ADD-style tracking reward after reference-percentile calibration. The authors derive the population-optimal identity J* = t/(1-t)^2 Cov[x|z] (Eq. 11, Appendix A) and use it to motivate PGS as a projection of the joint-change direction onto high-variance conditional co-variation modes. Experiments cover nine single-trajectory MimicKit skills, a 34-sequence general-tracking benchmark, ablations on prior scale, prediction parameterization, and reward formulation, and real-robot deployment with BeyondMimic. The paper claims improved sample efficiency and tracking accuracy, especially for dynamic motions.","tokens_in":19478,"tokens_out":8318,"duration_ms":82446,"significance":"If the central interpretation is accepted, the contribution is meaningful: a controller-independent pose prior that provides a cheap (one Jacobian-vector product), frozen, scaling-friendly reward signal for dynamic humanoid tracking, including tasks where vanilla ADD fails, with a plausible real-robot deployment. The strength of the paper is that PGS is derived rather than fitted: Eq. (11) and Eq. (16) follow from the conditional-mean regression objective, and the Appendix A derivation is verifiable step by step. The robustness study (Appendix B) usefully shows internal consistency and magnitude invariance. However, the paper's own Appendix A and Limitations statements explicitly identify the main approximation—agreement of the learned Jacobian with the population-optimal Jacobian—as unquantified, and this is the central load-bearing assumption for the geometric interpretation.","major_comments":[{"comment":"The geometric interpretation of PGS relies on J_phi(z,t_eval) approximating the population-optimal Jacobian J* at online rollout queries. Eq. (16) is proved for J*, but the method actually uses the Jacobian of a 10-block residual MLP. Appendix A itself concludes 'the interpretation is approximate to the extent that its input Jacobian agrees with the population-optimal Jacobian at the queried point,' and Appendix B explicitly notes that its perturbation study uses transitions drawn from the prior's training corpus and measures internal consistency only. No experiment quantifies agreement with the conditional covariance C_t(z) at held-out or online queries. If J_phi does not track the true conditional covariance, PGS is an arbitrary smoothness regularizer and the central claim is unsupported. Please add a quantitative faithfulness check, e.g., comparing J_phi d to an empirical covariance-v","section":"III-C, Eq. (11)-(16), Appendix A"},{"comment":"The evaluation timestep t_eval is selected by a validation sweep on the same 34-sequence LaFAN1 subset that is later used to report general-motion tracking results (Fig. 3). The reward-shaping hyperparameters p_good, p_bad, and alpha are set without a reported sensitivity analysis. This means the headline general-tracking gains may partly reflect selection on the evaluation benchmark. The robustness table (Table V) shows neighboring t_eval values behave similarly, which mitigates the concern, but final error metrics are only reported at t_eval=0.75. Please report the main general-tracking metrics across the t_eval sweep, or use a held-out validation split and show that the selected value transfers.","section":"III-D, Fig. 3, Table V"},{"comment":"PGS is sign-invariant at a fixed query because it is quadratic in d_k. The paper acknowledges in Limitations that 'opposite transitions may receive similar geometric evaluations' and that direction/order cannot be assessed. This substantially weakens the interpretation in Section I that PGS measures alignment with directional joint co-variation: a reversal of a coordinated transition receives the same score. For acrobatic skills such as backflip vs frontflip, direction is decisive. PGS may still serve as a subspace-alignment regularizer, and the signed tracking reward supplies the direction, but the mechanistic language should be revised (or a sign-sensitive score introduced) to avoid overstating what the prior reveals.","section":"VI Limitations / Eq. (15)"},{"comment":"The abstract's 'improves tracking accuracy' is not equally supported across all metrics. On single-trajectory tasks, the positional-error margins over PDF-HR are often within one standard deviation (e.g., Cartwheel 0.028±0.003 vs 0.028±0.003; Sideflip 0.049±0.001 vs 0.050±0.003; Speed Vault 0.021±0.003 vs 0.023±0.000). In general tracking, rotation errors overlap heavily (10 s: 0.119±0.009 vs 0.121±0.007). With only three seeds and no significance testing, the accuracy claim should be limited to the tasks and metrics where the gain is clear (sample efficiency, solving previously-failed Backflip and Double Kong) or supported by confidence intervals or effect sizes.","section":"IV-B Table II; Appendix D Table IX"}],"minor_comments":[{"comment":"The stabilization constant is written epsilon_num in the main paper and eta_num in Appendix A. Use a single symbol throughout.","section":"Eq. (14) / Appendix A"},{"comment":"Formatting is broken in the position-error rows for Backflip and Double Kong: 'Failed0.048' and 'Failed0.031' should be separated into 'Failed' and the numeric value.","section":"Table II"},{"comment":"The caption says 'min-normalized' without defining the normalization. State how the metrics are scaled and over which set of tasks.","section":"Fig. 2"},{"comment":"The text notes that PDF-HR yields errors comparable to or higher than vanilla ADD in general tracking, while Table II shows PDF-HR improves some single-trajectory tasks. A sentence explaining why the pose-distance field transfers less successfully to the multi-motion setting would be useful.","section":"IV-C"},{"comment":"Several references are arXiv preprints with future-dated numbers (e.g., 2608.03227 itself, 2602.04851, 2603.03279, 2603.27756). Please verify that all citation numbers and years are correct and add DOIs or URLs where available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central derivation is sound, and the experimental setup is appropriate. My major_revision recommendation is driven by the gap between the derived population-optimal PGS and the learned Jacobian actually used; the authors already hint at this gap in Appendix A, so I expect it can be addressed with an additional quantitative study rather than a conceptual rewrite. The sign-invariance and hyperparameter-tuning concerns are real but secondary, and they are partially acknowledged in the Limitations section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new thing is using a frozen flow-matching denoiser's Jacobian as a directional reward signal for RL pose tracking. The identity in Eq. (11) is known, but pointing out that you can score the rollout's joint-change direction with a Jacobian-vector product and calibrate it per reference is a real, practical contribution. The paper does good work: the derivation in Appendix A is correct, the ablations (FM-Recon vs. PGS, v-pred vs. x-pred, data scale) are informative, and the experimental evidence is consistent and includes a real-robot deployment. The gains on acrobatic tasks and the two tasks where vanilla ADD fails (Backflip, Double Kong) are believable, and the 0.75ms per-step cost is a strong practical point.\n\nThe soft spots, in order. First, the central interpretation: PGS is sold as measuring alignment with \"co-variation modes\" of the pose prior, but that interpretation requires the learned Jacobian J_phi to approximate the population-optimal J* at the online query points. Appendix A says this is approximate \"to the extent\" the Jacobians agree, but the paper never quantifies that extent. Appendix B checks internal consistency (discriminability, magnitude-invariance, rank correlation against a 128-sample oracle), but those checks would pass for any smooth network whose Jacobian is roughly oriented; they do not show J_phi is close to the true conditional covariance. So the mechanism could partly just be a smoothness regularizer. That doesn't kill the paper—the empirical gains stand on their own—but it does weaken the conceptual claim.\n\nThe second issue is the t_eval selection: the timestep is chosen by a validation sweep on the same 34-sequence LaFAN1 set used for the headline general-tracking numbers. That makes the general-tracking results less clean, though the neighboring timesteps (0.7, 0.8) perform nearly as well, so the tuning isn't knife-edged. Third, a minor point: no code or data release, which limits reproducibility. I also note the paper itself acknowledges the sign-invariance limitation for temporal transitions; that is honest and appropriate.\n\nOverall: the central method is sound enough, the experiments support the main empirical claims, and the weaknesses are addressable. I would send this to peer review with a request to see a faithfulness check of J_phi vs. J* and an evaluation of t_eval on held-out tasks. This is a useful paper for anyone working on physics-based character control or reusable motion priors; it will also be of interest to the flow-matching community as an application of denoiser geometry.","headline":"A frozen flow-matching pose prior used as a directional geometry reward is a genuinely new and useful idea; the main weakness is that the geometric interpretation rests on an unverified approximation of the learned Jacobian to the population-optimal one.","tokens_in":20033,"tokens_out":677,"would_cite":true,"duration_ms":8390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PFM-HR claims that scoring a rollout's joint-change pattern against the Jacobian of a flow-matching denoiser trained on unordered poses improves humanoid tracking, especially acrobatic motions.","keywords":["flow matching","pose prior","humanoid robots","reinforcement learning","motion tracking","pose geometry score","denoiser Jacobian","reward modulation"],"falsifier":"At online queries $\\tilde{q} = t_{\\mathrm{eval}} q + (1-t_{\\mathrm{eval}})\\varepsilon$, estimate the conditional covariance $C = \\mathrm{Cov}[x \\mid z = \\tilde{q}]$ by importance-weighted Monte Carlo over the pose corpus (weights proportional to $\\exp(-\\lVert \\tilde{q} - t_{\\mathrm{eval}} x \\rVert^2 / 2(1-t_{\\mathrm{eval}})^2)$), then compare the leading eigendirections of the learned Jacobian product $J^\\top J$ with those of $\\left(\\frac{t_{\\mathrm{eval}}}{(1-t_{\\mathrm{eval}})^2}\\right)^2 C^2$. If the average eigenvector alignment is no better than random, or if replacing the neural PGS with","tokens_in":19020,"feed_emoji":"🤖","tokens_out":17720,"duration_ms":158799,"temperature":0.7,"pith_summary":"PFM-HR sets out to show that a flow-matching denoiser trained only on unordered humanoid poses — no timestamps, no ordered motion clips, no pose-distance labels — still encodes how joints co-vary, and that this local geometry can be turned into a dense reward signal for reinforcement-learned tracking. The signal is the Pose Geometry Score: take the normalized joint-coordinate change between consecutive rollout poses, push it through the denoiser's Jacobian, and measure the squared response; a high score means the policy's pose transition aligns with strong co-variation modes of real poses. The score is mapped to a reference percentile and used to attenuate the tracking reward for unsupported change patterns while the prior stays frozen. Experiments report better sample efficiency and tracking accuracy than an adversarial-motion-prior backbone and a frozen pose-distance prior, including solving backflip and double-kong tracking that the backbone fails and cutting real-robot training samples by 15–31 percent.","feed_headline":"No clips needed: pose prior speeds humanoid acrobatics","feed_subtitle":"Scoring joint-change geometry against a frozen pose prior cuts training samples by up to 31 percent.","key_machinery":"The Pose Geometry Score (PGS), carried by the Jacobian–conditional covariance identity of the flow-matching clean-pose predictor. Because the denoiser is trained on the marginal pose distribution, its input Jacobian $J_\\varphi$ approximates the population-optimal map $\\nabla_z \\mathbb{E}[x \\mid z_t = z] = \\frac{t}{(1-t)^2}\\,\\mathrm{Cov}[x \\mid z_t = z]$; the score $s_{\\mathrm{PGS}} = \\lVert J_\\varphi d \\rVert_2^2$ is the squared response of that Jacobian to the normalized joint-coordinate change direction $d$ of a rollout transition. This makes PGS a reader of the local direction-pattern geometry of poses, computable as a single Jacobian–vector product; after mapping to the reference motion'","core_discovery":"PFM-HR's central discovery is that the input Jacobian of a clean-pose-predicting flow-matching network trained only on the marginal pose distribution encodes local joint co-variation geometry. For the population-optimal denoiser, $\\nabla_z \\mathbb{E}[x \\mid z_t = z] = \\frac{t}{(1-t)^2}\\,\\mathrm{Cov}[x \\mid z_t = z]$; applying the learned Jacobian to the normalized joint-coordinate change of a rollout step and squaring the norm gives the Pose Geometry Score — sensitive to the inter-joint pattern of a pose change, insensitive to its overall magnitude and global sign. A high score means the transition aligns with strongly represented co-variation modes of the pose prior; the tracking reward is","pith_inferences":["If the Jacobian-geometry reading is right, the recipe transfers: any RL problem with a static state-data manifold — dexterous manipulation, loco-manipulation — could score the policy's state-space update direction against a frozen generative prior's Jacobian, without needing temporal clips.","The paper's own limitation — PGS is sign-invariant and blind to transition direction and ordering, so a motion and its exact reverse can score alike — points to a natural next step that the authors also flag: a sign-sensitive or temporally conditioned variant that distinguishes direction, likely preserving the efficiency of the current score.","The ablation comparing PGS with a reconstruction-based score (FM-Recon) suggests the benefit comes from directional Jacobian response rather than pose plausibility per se; a cheaper surrogate for tangent-manifold alignment might achieve similar gains in other settings.","Because the single-noise-sample estimator already agrees with a 128-sample reference on 97.9% of calibration regions, the practical variance of the score is small; the open question the paper leaves is how close the learned Jacobian is to the true conditional covariance at online queries — a diagnostic that would settle whether the guidance is genuinely geometric or a smoothness bias."],"forward_implications":["A pose prior becomes a plug-in: one frozen model, trained once on 60M unordered poses, can be attached to different tracking policies and tasks with no retraining and no ordered data.","Scale is cheap: growing the prior corpus from 30M to 60M poses costs 65 GPU-hours by continued training, versus 100 from scratch and over 500 for the pose-distance supervision the PDF-HR baseline needs.","Dynamic skills gain most: on Backflip and Double Kong the vanilla backbone never converges, but both prior-based variants solve them and PFM-HR converges fastest.","In general motion tracking over 10/20/30 s horizons, averaged position error falls 7.6% versus the ADD backbone and 10.3% versus PDF-HR, with rotation error down 3.6% and 7.3%.","The reward is cheap: one Jacobian–vector product per step (0.75 ms for a batch of 4096) versus 1.8 ms for three reconstruction evaluations, so dense geometric guidance adds little overhead to RL."],"supporting_citations":[{"why":"Supplies the linear flow-matching path $z_t = t x + (1-t)\\epsilon$ and the training objective for the pose prior.","marker":"[11]"},{"why":"Provides the Jacobian–conditional covariance identity $\\nabla_z \\mathbb{E}[x|z] = \\frac{t}{(1-t)^2}\\mathrm{Cov}[x|z]$ that grounds the PGS geometric interpretation.","marker":"[19]"},{"why":"The clean-pose (x-prediction) parameterization (JiT) whose input Jacobian is exactly what PGS reads.","marker":"[9]"},{"why":"The adversarial differential discriminator backbone (ADD) that PFM-HR augments and all compared baselines share.","marker":"[38]"},{"why":"The frozen pose-distance-field prior (PDF-HR) that is the main controlled baseline and the strongest prior-only alternative.","marker":"[5]"},{"why":"The 60M-pose BONES-SEED corpus used for prior pretraining and the nested-scale ablation.","marker":"[2]"},{"why":"The LaFAN1 motion corpus whose 34-sequence subset is the general-motion tracking benchmark.","marker":"[6]"},{"why":"The BeyondMimic training and deployment pipeline used to show PFM-HR's simulation-only gain transfers to real-robot deployment.","marker":"[10]"},{"why":"The MimicKit framework providing the nine single-motion tracking tasks and the official ADD training configuration.","marker":"[23]"}],"fun_headline_variants":["Flow matching on unordered poses speeds humanoid tracking","Geometry score guides humanoid pose transitions","Frozen pose prior cuts training samples by up to 31 percent","No clips needed: pose-based prior for humanoid acrobatics"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the finite-capacity denoiser's Jacobian, evaluated at the stochastically corrupted rollout queries, behaves like the population-optimal conditional covariance of clean poses — a faithfulness the paper derives in theory but never measures directly at those query points.","fun_headline_variants_meta":{"raw":{"variants":["Flow matching on unordered poses speeds humanoid tracking","Geometry score guides humanoid pose transitions","Frozen pose prior cuts training samples by up to 31 percent","No clips needed: pose-based prior for humanoid acrobatics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":970,"prompt_tokens":661,"completion_tokens":309,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":405,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":405,"tokens_out":309,"duration_ms":3767,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:05:33.357519+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At online queries $\\tilde{q} = t_{\\mathrm{eval}} q + (1-t_{\\mathrm{eval}})\\varepsilon$, estimate the conditional covariance $C = \\mathrm{Cov}[x \\mid z = \\tilde{q}]$ by importance-weighted Monte Carlo over the pose corpus (weights proportional to $\\exp(-\\lVert \\tilde{q} - t_{\\mathrm{eval}} x \\rVert^2 / 2(1-t_{\\mathrm{eval}})^2)$), then compare the leading eigendirections of the learned Jacobian product $J^\\top J$ with those of $\\left(\\frac{t_{\\mathrm{eval}}}{(1-t_{\\mathrm{eval}})^2}\\right)^2 C^2$. If the average eigenvector alignment is no better than random, or if replacing the neural PGS with","supporting_citations":[{"cited_title":"Physics-based motion imitation with adversarial differential discriminators","cited_arxiv_id":null,"evidence_quote":"The adversarial differential discriminator backbone (ADD) that PFM-HR augments and all compared baselines share."},{"cited_title":"Pdf-hr: Pose distance fields for humanoid robots.arXiv preprint arXiv:2602.04851, 2026","cited_arxiv_id":null,"evidence_quote":"The frozen pose-distance-field prior (PDF-HR) that is the main controlled baseline and the strongest prior-only alternative."},{"cited_title":"Ai datasets for machine learning and motion capture","cited_arxiv_id":null,"evidence_quote":"The 60M-pose BONES-SEED corpus used for prior pretraining and the nested-scale ablation."},{"cited_title":"Robust motion in-betweening.ACM Transactions on Graphics (TOG), 39(4):60–1, 2020","cited_arxiv_id":null,"evidence_quote":"The LaFAN1 motion corpus whose 34-sequence subset is the general-motion tracking benchmark."}],"review_version":1}