{"id":"76d09ccf-45bb-4f3f-bbcf-831b127cd5f9","arxiv_id":"2607.10984","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"A permutation-equivariant latent diffusion model treats skeleton connectivity as input, enabling the first kinematics-agnostic stochastic human motion predictor that generalizes zero-shot to unseen and partial skeletons.","lead":"EquiFusion is a latent diffusion model for stochastic 3D human motion prediction that stays agnostic to skeleton joint count and connectivity by making the adjacency matrix an explicit input and enforcing joint-order permutation equivariance. This removes the need for per-skeleton networks or error-prone retargeting and enables zero-shot prediction on unseen or partial skeletons while matching or beating prior accuracy with far fewer parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified that undermines the central kinematics-agnostic claim.","rationale":"The reader's strongest claim accurately captures the paper's contribution and evidence. The identified weakest assumption is precisely the softest remaining point (PEQ + A guarantees order/size independence and enables the architecture, but structure generalization on non-homeomorphic graphs is ultimately inductive and empirically demonstrated rather than purely mathematical). Because the ablations, proofs, multi-dataset experiments, and retargeting-controlled baselines already address it adequately, and because the acknowledged limitations (non-leaf bones, modest data-volume effects) are peripheral, no adjustment to the ACCEPT verdict is warranted. The work is solid, reproducible (code promised), and advances a genuine bottleneck.","tokens_in":51631,"tokens_out":564,"duration_ms":22192,"concrete_test":"Numerically re-verify end-to-end equivariance (fixed-noise case of Eq. 32) on a held-out non-homeomorphic skeleton never seen in training (e.g., FreeMan ears or a synthetic 15-joint chain with novel end-effectors): confirm ||f(PX, P A P^T, P ε) - P f(X, A, ε)|| < 1e-5 and that zero-shot uADE remains within 10% of the H36M numbers in Tab. 1. If both hold, the sufficiency claim is reinforced; if either fails, the design guarantee is narrower than stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core argument (Lemma 1 + Theorem 2 + end-to-end PEQ via Eqs. 3-4 and proofs in C.2, with A as explicit input) correctly establishes that shared weights independent of J and order are necessary for cardinality-agnosticism and sufficient for the claimed design property. Ablations (Tab. 4: failure of 'Ours w/o Eq' and of joint-dependent multi-dataset SkelDiff variant; success of 'Ours on P') plus zero-shot tables (Tabs. 1, 3, 11) and constant-parameter scaling (Fig. 3) supply consistent empirical support. The residual soft spot noted by the reader (non-homeomorphic generalization without joint semantics; non-leaf occlusion handling via bone directions) is real but secondary: semantics add nothing (Ours+S), multi-dataset gains are small relative to architecture, and the paper already flags the leaf-only limitation. These do not falsify the design-level claim or the SOTA/efficiency results.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes EquiFusion, claimed as the first kinematics-agnostic model for stochastic 3D human motion prediction (SHMP). It implements a latent diffusion model whose autoencoder and denoiser are end-to-end permutation-equivariant with respect to joint ordering (Eq. 2), taking the skeleton adjacency matrix A as an explicit input. This design satisfies the parameter-cardinality independence of Lemma 1 via Theorem 2, enabling a single model to train on heterogeneous kinematics (AMASS + Nymeria) and to perform zero-shot inference on unseen full-body skeletons (H36M) and partial/occluded limbs without retargeting or ad-hoc occlusion training. Empirically it reports state-of-the-art or competitive precision, diversity and realism on AMASS, H36M and Nymeria while using up to 75 % fewer parameters than the strongest kinematics-specific diffusion baseline and supporting constant scaling with the number of kinematics.","tokens_in":51881,"tokens_out":1066,"duration_ms":19598,"significance":"If the claims hold, the work removes a long-standing structural bottleneck in SHMP: hard-coded skeleton graphs that force per-dataset models, expensive retargeting, and inability to handle occlusions or novel sensors. The combination of a clean linear-algebra equivariance argument (Secs. C.1–C.2), an explicit adjacency input, bone-direction motion representation that guarantees limb-length consistency by construction, and extensive zero-shot and multi-dataset experiments constitutes a genuine advance toward foundation-style human-motion models. Concrete strengths include machine-checkable equivariance proofs, public code, constant-parameter scaling (Fig. 3), and ablations that isolate the contribution of equivariance (Tab. 4). These results are immediately useful for robotics, AR and multi-sensor pipelines that must operate across heterogeneous MoCap formats.","major_comments":[{"comment":"Sec. 3.2–3.3 and Lemma 1 / Theorem 2 correctly establish that joint-order permutation equivariance plus an explicit adjacency matrix yields cardinality-independent parameters. The residual claim that this is already sufficient for true non-homeomorphic kinematics-agnosticism (different end-effectors, different joint counts) without joint-type semantics rests primarily on the empirical zero-shot tables (Tabs. 1, 3, 11) and the “Ours+S” ablation. While the ablation shows no gain from semantics, a short additional experiment that systematically varies end-effector sets (or reports failure cases on more distant graphs) would strengthen the sufficiency argument beyond the current human-only skeletons.","section":null},{"comment":"Sec. 5 Limitations correctly notes that bone-direction decoding handles only leaf missing joints. Because partial-kinematics zero-shot is advertised as a core novel capability (Abstract, Sec. 1, Tab. 3), the paper should either (a) quantify the fraction of real-world occlusions that are leaf-only or (b) supply a simple non-leaf completion baseline so that readers can judge the practical scope of the claimed occlusion handling.","section":null}],"minor_comments":[{"comment":"Fig. 3 (left) would be clearer if the y-axis were log-scaled or if absolute parameter counts for each baseline under multi-kinematics were listed in a small table; the visual claim of constant scaling is important and currently a bit hard to read.","section":null},{"comment":"Unified metrics (uADE etc.) are a welcome contribution (Sec. 4.1, F.1). A one-sentence reminder in every table caption that conventional ADE ranks identically would help readers who still compare against older papers that report only ADE.","section":null},{"comment":"Sec. D.3 and Fig. 5: the forward-kinematics rescaling step that restores bone lengths from the observation is described clearly, yet a short pseudocode block would make the pipeline fully reproducible without consulting the (promised) code.","section":null},{"comment":"Occasional typographical inconsistencies appear (e.g., “i.e.kinematics”, missing spaces after commas in several places). A final proof-reading pass would polish the manuscript.","section":null},{"comment":"Tab. 4 “Ours on P+Pϵ” versus “Ours on P”: the distinction between sample-wise and distributional equivariance is correctly explained in Sec. C.3; a single clarifying sentence in the table caption would prevent mis-reading by non-specialists.","section":null}],"recommendation":"accept","confidential_remarks":"The work is solid, well-executed and timely. The only soft spot is the precise scope of “kinematics-agnostic” for non-homeomorphic graphs; the authors already flag the leaf-only limitation, so I do not view it as blocking. Code release and clean proofs raise the bar for the sub-field. Suitable for a top venue after the two minor clarifications above."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The core advance is real: EquiFusion is the first SHMP model that treats the adjacency matrix as an explicit input and enforces end-to-end permutation equivariance over joints (Eq. 2, graph conv + GAT in Eqs. 3-4, proofs in C.2). That design choice, plus the bone-direction parametrization, lets one network train across AMASS+Nymeria and run zero-shot on H36M kinematics or random limb occlusions without retargeting or retraining. Lemma 1 / Theorem 2 correctly tie parameter independence of J to equivariance; the ablations (Tab. 4: breaking equivariance or using joint-dependent multi-dataset SkelDiff both fail zero-shot) and constant-parameter scaling (Fig. 3) back it up. SOTA-competitive numbers on the usual single-kinematics splits, 75% fewer parameters than SkelDiff, and faster train/infer are earned, not claimed.\n\nWhat works well is the experimental hygiene: fair (if imperfect) Holden retargeting baselines for the competitors, unified metrics that let you compare across joint counts, explicit checks that the model stays equivariant under joint/adjacency/noise permutations, and code release. Multi-dataset training actually helps a bit on the harder zero-shot tables, and body-realism is perfect by construction.\n\nSoft spots are secondary and already flagged by the authors. Bone-direction decoding only cleanly handles leaf missing joints; non-leaf occlusions need the adjacency mask plus a heuristic. Multi-dataset gains are partly data volume, not pure topology magic. Semantics add nothing (Ours+S). Retargeting error cannot be fully isolated, but the triangle-inequality upper bound and the fact that EquiFusion still wins by a clear margin keep the comparison honest. None of this undercuts the design claim or the practical utility for robotics/AR pipelines that already face heterogeneous MoCap or partial observations.\n\nThis is for anyone building motion models that must leave the single-skeleton sandbox. The math is clean linear algebra, the data are public held-out splits, citations are appropriate. I would bring it to reading group, cite the equivariance construction and the zero-shot tables, and send it to peer review without hesitation.","headline":"Clean first kinematics-agnostic SHMP via end-to-end joint-order equivariant latent diffusion; zero-shot full/partial skeletons and multi-dataset training work, with real efficiency gains and solid proofs.","tokens_in":52482,"tokens_out":578,"would_cite":true,"duration_ms":13005,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"One equivariant network predicts future human motion for any skeleton layout, including ones never seen in training.","keywords":["stochastic human motion prediction","kinematics-agnostic","permutation equivariance","latent diffusion","zero-shot kinematics","cross-dataset training","bone directions","occlusion handling"],"falsifier":"Train only on AMASS (22 joints), then evaluate zero-shot on H36M (17 joints) and on randomly occluded AMASS limbs; if the unified ADE, FID and body-realism gaps versus the best retargeted specialised baselines reverse or vanish, the claim fails.","tokens_in":52554,"feed_emoji":"🦴","tokens_out":607,"duration_ms":5446,"temperature":0.7,"pith_summary":"Stochastic human motion prediction has long been locked to a single skeleton: every new joint set or capture system forced a fresh network, retargeting errors, and no way to train across datasets. EquiFusion removes that lock. It treats the skeleton’s connectivity matrix as an ordinary input and builds every layer so that permuting the joint order simply permutes the output. The resulting latent diffusion model can therefore be trained once on one or several skeletons and still produce diverse, realistic futures for entirely new full-body layouts, for skeletons with missing limbs, and for out-of-distribution motions. On standard benchmarks it matches or beats the previous best specialised models while using up to 75 percent fewer parameters and running faster. The practical payoff is a single, compact predictor that works across capture systems and occlusion patterns without engineering a new architecture each time.","feed_headline":"One network predicts motion for any human skeleton","feed_subtitle":"Equivariant diffusion generalises zero-shot to unseen joints and missing limbs, 75% smaller","key_machinery":"End-to-end joint-order permutation equivariance (Lemma 1 / Theorem 2 / Eq. 2): every graph convolution and attention layer is written so that reordering the joints and the adjacency matrix simply reorders the output, guaranteeing that the number of parameters is independent of the number of joints.","core_discovery":"Permutation equivariance with respect to joint order, together with an explicit adjacency-matrix input, is sufficient to make a latent diffusion model kinematics-agnostic: a single set of weights can train on heterogeneous skeletons and generalise zero-shot to unseen full-body and partial kinematics while remaining competitive on conventional single-skeleton benchmarks.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["One model predicts 3D motion for any human skeleton","Equivariant diffusion forecasts motion across unseen kinematics","Kinematics-agnostic network handles zero-shot partial observations","Permutation equivariance enables cross-skeleton motion prediction","Single weights generalise to heterogeneous human skeletons"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"That making the network ignore joint order and take connectivity as input is enough for true generalisation across skeletons that have different end-effectors and different numbers of joints, without any joint-type labels or other extra bias.","fun_headline_variants_meta":{"raw":{"variants":["One model predicts 3D motion for any human skeleton","Equivariant diffusion forecasts motion across unseen kinematics","Kinematics-agnostic network handles zero-shot partial observations","Permutation equivariance enables cross-skeleton motion prediction","Single weights generalise to heterogeneous human skeletons"]},"model":"grok-4.5","effort":"low","cost_usd":0.00787,"raw_usage":{"total_tokens":1861,"prompt_tokens":728,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":78700000,"prompt_tokens_details":{"text_tokens":728,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1055,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":728,"tokens_out":78,"duration_ms":10928,"temperature":1.0,"reasoning_tokens":1055,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T07:51:18.283177+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train only on AMASS (22 joints), then evaluate zero-shot on H36M (17 joints) and on randomly occluded AMASS limbs; if the unified ADE, FID and body-realism gaps versus the best retargeted specialised baselines reverse or vanish, the claim fails.","supporting_citations":[],"review_version":1}