{"id":"67ea94cd-96c6-4724-b905-37dc1140b5c3","arxiv_id":"2608.03158","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Surface keypoint trajectories plus a spatio-temporal contact distance field let one generative model synthesize human motion for single, multiple, and articulated objects without specifying joint types.","lead":"To make virtual people interact naturally with furniture, boxes, or drawers, this paper tracks a few surface points on each object or object part and generates motions for both. It also learns a contact field that says which body region should touch which object at each moment, enabling a single model to handle one object, several objects, and articulated objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The joint-type-agnostic claim is only approximately supported: Appendix Table 6 reports 5.20% assembly-connectivity violations and 1.46% axis-deviation frames, and this residual is untested at longer horizons or stricter thresholds.","rationale":"The reader's weakest assumption pins the central claim on the model's ability to infer articulation mechanisms from keypoint trajectories without explicit joint constraints. The paper's own kinematic audit supplies direct evidence that this assumption is only approximately met: nonzero hinge residuals in Table 3 and 5.20% assembly-connectivity violations in Appendix Table 6. I agree this is the load-bearing soft spot. The concern does not force rejection: the reported violation rates are low, the paper discloses them, and an optional correction step is available. But it does justify a conditional verdict, because the current evaluation does not show whether the residual is bounded or grows with sequence length. A horizon test is the cleanest way to distinguish 'learned articulation mechanism' from 'short-horizon trajectory statistics.' If drift accumulates, the central contribution of joint-type-agnostic articulated-object generation is materially weakened, and the method would need an explicit kinematic prior or constraint, which is exactly what it claims to avoid. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":25940,"tokens_out":9342,"duration_ms":87121,"concrete_test":"Run the trained Stage I object-motion generator (without the optional projection correction) on ParaHome articulated-object test sequences at T=240 and T=480, with training/validation horizon T=120, and recompute the Appendix Table 6 assembly-connectivity violation rate using both the original 2 cm threshold and a stricter 1 cm threshold. If the violation rate increases with T (e.g., exceeds 10% at T=240) or the mean pivot drift surpasses 1 cm, the model is producing locally plausible but globally drifting trajectories rather than maintaining a stable articulation prior, and the joint-type-agnostic claim fails for longer interactions. As a secondary check, measure the per-frame pairwise-distance drift among each generated component's three keypoints; if the mean relative drift exceeds roughly 5% of object scale, the Kabsch recovery in Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. 3.1 is that surface keypoint trajectories are a lossless, joint-type-agnostic representation because different articulation mechanisms produce distinctive point trajectory patterns that a generative model can learn directly from data, without explicit joint-type specification. For that claim to hold, the Stage I generator must implicitly enforce that a movable part stays on its joint axis and remains assembled to its base; nothing in the architecture (Eq. 2) or the keypoint/Kabsch decoding (Eq. 1) imposes these constraints. The paper's own kinematic audit (Appendix Table 6, 282 part-base pairs) reports 1.46% frames with joint-axis deviation above 10 degrees, 0.02% joint-limit violations, and 5.20% assembly-connectivity violations at a 2 cm threshold, with mean pivot drift 0.43 cm. These are nonzero residuals and show the premise is only approximately satisfied. The audit is reported only at the training horizon T=120; it does not establish that the model has learned a stable articulation mechanism rather than locally plausible trajectory statistics. If violations accumulate with sequence length, the 'without explicit joint-type specification' contribution is weakened, and the optional 'Ours + Correction' step in Table 3, which requires a predefined pivot/axis and joint type, would become necessary. A related untested assumption is intra-component rigidity: because Eq. (1) recovers pose from three non-collinear keypoints, any drift in their pairwise distances makes Kabsch fit a non-rigid point set, so the 'lossless' property is not verified on generated outputs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces surface keypoint trajectories as an alternative to SE(3) object representations for human-object interaction generation. Each rigid component is represented by K non-collinear surface keypoints, and the authors use the Kabsch algorithm to recover its pose. Contact is modeled by a sigmoid contact distance field between 138 SMPL-X body markers and 384 surface points per component. The proposed method factorizes generation into three diffusion stages: object keypoint trajectory generation, contact distance field prediction, and human marker motion synthesis with diffusion noise optimization for contact refinement. Experiments are reported on ParaHome, HIMO, ARCTIC, and OMOMO, with comparisons against HIMO-Gen, CoDA, HOI-Diff, CHOIS, ROG, and HOIDiNi, together with ablations, a perceptual user study, and kinematic-validity audits.","tokens_in":26293,"tokens_out":6904,"duration_ms":62739,"significance":"If the central claims hold, the keypoint-trajectory representation would be a useful unified object representation that avoids SE(3) manifold issues and explicit joint-type modeling, and the spatio-temporal contact distance field is a richer supervision signal than binary contact labels. The paper is strong in scope: it evaluates on four benchmarks, reports ablations for the object representation and the contact formulation, conducts a perceptual user study, and provides reconstruction and kinematic-validity audits. The strengths include the explicit robustness analysis for Kabsch recovery, the permutation study for slot ordering, and the optional projection step that makes residual articulation errors transparent. However, the load-bearing articulation and contact-accuracy claims need qualification and more stringent evaluation before the paper's central claims can be taken at face value.","major_comments":[{"comment":"The paper's central claim that articulated objects are generated 'without explicit joint-type specification' is only approximately supported by the reported evidence. Since Eq. (1) recovers component pose from K non-collinear keypoints and the Stage I diffusion model (Eq. (2)) imposes no joint constraints, generated components are not guaranteed to stay on their joint axes or remain assembled. The paper's own audit (Appendix Table 6) reports 5.20% frame-level assembly-connectivity violations at a 2 cm threshold and 1.46% frames with axis deviation above 10 degrees, and the optional 'Ours + Correction' (Appendix Table 3) requires a predefined pivot/axis and joint type. Please state explicitly whether any main-table result uses the correction, report how these residuals behave beyond the T=120 training horizon, and clarify when the correction is necessary; without that, the 'joint-type-agnostic' contribution is considerably weaker than claimed.","section":"Sec. 3.1, Appendix Table 6, Appendix Table 3"},{"comment":"The contact accuracy metric is structurally aligned with both the contact field definition and the Stage III optimization target, which can inflate the reported contact improvements. Eq. (3) defines the field as sigmoid((tau - dist)/alpha), Appendix Eq. (9) scores contact with the same thresholded-distance notion at tau=2 cm, and Eq. (6) optimizes L_contact against the predicted field by thresholding D_hat. Thus, part of the improvement in C_tem/body acc may reflect agreement of human markers with the model's own predicted field rather than with genuinely observed contact. I recommend reporting contact accuracy under alternative thresholds (e.g., 1 cm and 3 cm) or a surface-area/contact-patch metric that is not derived from Eq. (3), and reporting the no-DNO ablation in Table 1 with the same metrics to quantify how much of the gain comes from optimizing the measured quantity.","section":"Eq. (3), Eq. (6), Appendix Eq. (9)"},{"comment":"Most head-to-head numbers are reported for a single run without error bars, and several differences are small. For example, Table 5 shows C_body acc of 0.902 vs 0.896 for the Euclidean-distance baseline and Table 1 shows 0.896 vs 0.902 across object counts; Table 2 is the only table with error bars. This makes it difficult to judge whether the claimed advantage over baselines and the ranking of ablations are statistically robust. Please provide at least three seeds (or a paired evaluation over the test set) and report mean plus/minus standard deviation or confidence intervals for the main tables, and state the significance test used.","section":"Tables 1, 3, 4, 8, 9"}],"minor_comments":[{"comment":"The Fig. 3 caption refers to the Stage II model as f_c, but Eq. (4) uses f_d; unify the notation.","section":"Fig. 3 vs Eq. (4)"},{"comment":"The Table 2 caption cites 'CoDA [Lv et al. 2024]', but CoDA is [Pi et al. 2025] in the text and references; fix the citation.","section":"Table 2 caption"},{"comment":"L_contact in Eq. (6) is not defined in the main text, and the Appendix sentence after Eq. (6) does not contain its explicit formula; define it in the main text or point to the exact supplement equation.","section":"Eq. (6)"},{"comment":"Table 4 reports T_obj and O_obj without defining them in the main text; define translation error and orientation error and state how they are computed from the recovered poses.","section":"Table 4"},{"comment":"The hyperparameters tau and alpha in Eq. (3), along with K, N, M_c, and Q, are only given in the appendix; add a single hyperparameter table in the main paper because the representation claims depend on these choices.","section":"Eq. (3) and Sec. 4.1"},{"comment":"The main-text and appendix table numbering collide (both contain a Table 3); renumber the appendix tables to avoid confusion.","section":"Appendix numbering"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about its residuals and optional correction, and the issues raised are fixable with additional experiments rather than being conceptual dead ends. Given the representation-centric claim, code and checkpoints would be valuable; their absence, combined with single-run tables, increases the burden on the quantitative claims. The citation pattern is fine, and I see no novelty-disclosure concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious referee. The core idea—representing each rigid component by a small set of surface keypoints and letting the generative model discover articulation patterns from trajectory geometry—is genuinely new relative to SE(3)-based baselines, and the three-stage factorization is clean. The authors do a lot right: four benchmarks, ablations on the representation and the contact field, a user study, and an appendix that audits kinematic validity rather than just showing pretty frames.\n\nThe honest caveat is that the headline claim is not fully settled. The surface-keypoint representation is lossless for rigid motion in theory, but nothing in the architecture enforces intra-component rigidity on generated outputs; the Kabsch recovery in Eq. 1 will fit a non-rigid point set if the three keypoints drift. The paper's own Table 6 shows 5.2% assembly-connectivity violations and 1.46% axis-deviation frames at 2cm and 10-degree thresholds. Those numbers are low, and the authors present them fairly, but they rule out the stronger reading of \"joint-type-agnostic\" as \"implicitly enforces kinematics.\" The optional correction step in Table 3, which requires a predefined pivot/axis and joint type, is effectively an admission that the agnostic generator alone is not exact. I don't think this sinks the paper—the residuals are small and the correction is a sensible engineering fallback—but the abstract oversells slightly.\n\nThe other soft spots are standard but worth flagging: no code release, most tables are single runs without error bars (ARCTIC has them, ParaHome/HIMO/OMOMO don't), and the contact accuracy metric in the appendix uses exactly the 2cm thresholded-distance notion that the model is trained against, so some of the reported contact gains are partially circular. That doesn't make the gains fake; the ablation against binary labels and raw Euclidean distance shows the field representation itself matters. But it does mean the absolute contact numbers are not independent evidence.\n\nWho is this for? Anyone working on HOI generation, especially articulated or multi-object settings. The representation is simple enough to adopt and the appendices contain enough detail to reproduce without the code. I'd engage with it and I'd cite it.\n\nRecommendation: send it to peer review. It deserves referee time, and a conditional accept that pushes for code release and a longer-horizon kinematic test would be about right.","headline":"A genuinely useful representational contribution to HOI generation with unusually honest kinematic auditing, but the joint-type-agnostic claim is only approximately true and the contact metric partially shares its objective with the training loss.","tokens_in":26824,"tokens_out":2458,"would_cite":true,"duration_ms":22632,"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":"Surface keypoint trajectories are a lossless, joint-type-agnostic representation of object motion that lets a single diffusion pipeline generate multi-object and articulated human-object interactions.","keywords":["human-object interaction generation","surface keypoint trajectories","articulated object motion","multi-object interaction","contact distance field","whole-body motion synthesis","diffusion models","motion representation"],"falsifier":"Prompt a trained model to open a prismatic drawer in a cabinet it has not seen, fit a line to each generated drawer trajectory, and measure the per-frame deviation of the drawer's recovered pose from that line and from the cabinet's track. If the median deviation exceeds the paper's own 2 cm assembly-connectivity threshold, or if the recovered joint value leaves the ground-truth range by more than 10% for a substantial fraction of frames, the claim that joint mechanisms are learned from keypoint trajectories alone is refuted; the paper's own appendix reports a 5.20% frame-level connectivity violation rate, so the test should show that rate stays near zero across novel articulated objects.","tokens_in":25751,"feed_emoji":"🖐️","tokens_out":12305,"duration_ms":99095,"temperature":0.7,"pith_summary":"The paper is trying to establish that a handful of non-collinear surface points per rigid part is a complete object motion representation for human-object interaction generation, and that this representation removes the usual need to pre-specify how many objects are involved or what kind of joint an articulated object has. Three points determine a rigid transformation in closed form, so the representation is lossless for rigid motion; and hinge, sliding, and screw joints each trace distinctive point trajectories, so a generative model can learn the articulation mechanism from data rather than from an explicit joint label. The same representation fits into a fixed-size slot buffer, so one model can generate interactions with a variable number of objects or with multi-part articulated objects. The paper adds a spatio-temporal contact distance field that records which of 138 body surface markers is near which object surface point at each frame, and a three-stage diffusion pipeline that first generates object motion, then the contact field, then body motion. If this holds, one text-conditioned model could cover single-object, multi-object, and articulated interactions without per-scene bookkeeping; across four benchmarks the authors report better or comparable motion quality and contact accuracy than $\\mathrm{SE}(3)$-based baselines.","feed_headline":"Three keypoints per part unlock one-model human-object generation","feed_subtitle":"A surface-keypoint representation handles multi-object and articulated scenes without joint-type labels.","key_machinery":"The load-bearing object is the surface keypoint trajectory. For each rigid component, $K$ canonical surface points are chosen by farthest point sampling ($K=3$ in the main experiments); at every time step their global 3D positions form the component's state, and the per-frame rigid pose is recovered in closed form by minimizing $\\|R O_0 + t - O_t\\|^2$ via the Kabsch algorithm, which is exact whenever the keypoints are non-collinear. Because the representation lives in Euclidean space, the generative model never has to predict on the $\\mathrm{SE}(3)$ manifold, and because unused slots in a fixed-size component buffer are masked during training, a single Transformer-based flow-matching diffusion model can handle any number of components. The second mechanism is the contact distance field $D_t = \\sigma((\\tau - \\mathrm{dist}(H_t^c, S_t))/\\alpha)$, computed between a subset of 47 body markers and 384 dense object surface points per component; Stage II predicts this field, and Stage III uses it as a soft contact prior together with an object signed-distance-field penetration penalty. Stage III performs Diffusion Noise Optimization, updating the initial noise rather than the output motion, so the refined motion stays on the learned manifold while contact violations are reduced.","core_discovery":"The central claim, stated on the paper's own terms, is that surface keypoint trajectories are a lossless and joint-type-agnostic object motion representation: for each rigid component, a small set of non-collinear surface points tracked over time fully determines the component's pose via the Kabsch algorithm, and the trajectory patterns of revolute, prismatic, and screw joints are distinctive enough for a generative model to learn them directly from data. The paper further claims that a spatio-temporal contact distance field, computed as a sigmoid of the distance between 138 body surface markers and 384 surface points per component, captures when and where each body region contacts each object, and that this field is a more learnable intermediate signal than binary contact labels, raw Euclidean distances, or fixed contact pairs. The two representations are combined in a factorized three-stage pipeline—object motion diffusion, contact-field diffusion, and body motion diffusion refined by contact-guided noise optimization—and the authors report that this pipeline matches or outperforms existing methods on single-object, multi-object, and articulated interaction benchmarks.","pith_inferences":["Editorial inference: because the representation is joint-type-agnostic, the same pipeline could plausibly extend to more complex mechanisms such as multi-axis linkages or gear trains without architectural changes; the paper does not test this.","Editorial inference: the paper's robustness table shows over-determined keypoint sets reduce rotation error under noise, while its generation ablation favors three keypoints; a natural hybrid is to generate with three keypoints but recover pose from a denser set, a combination the paper does not evaluate.","Editorial inference: the predicted contact distance field is a standalone, time-varying whole-body contact prior that could be reused for retargeting or physics-based control beyond the three-stage pipeline, which the paper leaves unexamined.","Editorial inference: the approximately four-minute per-sequence cost is dominated by noise optimization; a testable extension suggested by the paper's own discussion is to distill the optimized noise into a feed-forward correction network, which would make real-time deployment plausible."],"forward_implications":["A single generation model can produce interactions with one object, several objects, or an articulated assembly without being told the object count or the joint type first.","Object motion can be generated directly from text or sparse waypoints as point trajectories and converted to rigid poses afterward, so no separate per-object pose parameterization or per-joint handling is needed.","The spatio-temporal contact distance field provides a learnable whole-body contact signal; the paper's ablations find it improves contact accuracy over binary labels, raw Euclidean distances, and fixed contact pairs.","Articulated generation extends beyond revolute joints to prismatic and screw mechanisms; when exact joint constraints are required, the paper shows the generated trajectory can be projected onto the constrained motion without retraining.","On unseen object compositions, the method keeps object motion smooth and contact accuracy high relative to a fixed-count baseline, suggesting the representation generalizes to object combinations not seen together in training."],"supporting_citations":[{"why":"It gives the closed-form least-squares rotation and translation recovery that makes three non-collinear surface points a lossless rigid pose.","marker":"[Kabsch 1976]"},{"why":"It supplies farthest point sampling, used to choose the canonical surface keypoints on each rigid component.","marker":"[Qi et al. 2017]"},{"why":"It provides the SMPL-X body model from which the 138 surface markers are taken, placing human motion in Euclidean space.","marker":"[Pavlakos et al. 2019]"},{"why":"It supplies Basis Point Sets, which encode each object's canonical geometry as conditioning for the diffusion stages.","marker":"[Prokudin et al. 2019]"},{"why":"It provides Diffusion Noise Optimization, used in Stage III to enforce contacts while keeping the motion on the learned manifold.","marker":"[Karunratanakul et al. 2024]"},{"why":"ParaHome supplies the multi-object and articulated interaction sequences used for the main evaluation and for the unseen-composition test.","marker":"[Kim et al. 2025]"},{"why":"HIMO supplies a multi-object benchmark and the HIMO-Gen baseline; the paper's comparison shows a single variable-count model outperforms fixed-count multi-object models.","marker":"[Lv et al. 2024]"},{"why":"It is the closest prior contact distance field; the paper extends its single-object, skeleton-based field to whole-body multi-object settings and compares against it.","marker":"[Xue et al. 2025]"},{"why":"It is the articulated-object baseline CoDA that hardcodes a hinge axis, providing the contrast for the claim that surface keypoints handle revolute, prismatic, and screw joints without joint-type specification.","marker":"[Pi et al. 2025]"}],"fun_headline_variants":["Keypoint trajectories replace joint labels in human-object generation","Three keypoints per part enable multi-object articulated HOI synthesis","Surface keypoints make object motion joint-type agnostic","Multi-object and articulated HOI from surface keypoint dynamics","Contact distance fields guide whole-body motion from keypoint flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach rests on the assumption that a learned model can read the type of a joint—hinge, sliding, or screw—straight off the trajectories of a few surface points, so articulated parts stay on their axes and stay connected without any explicit joint constraint.","fun_headline_variants_meta":{"raw":{"variants":["Keypoint trajectories replace joint labels in human-object generation","Three keypoints per part enable multi-object articulated HOI synthesis","Surface keypoints make object motion joint-type agnostic","Multi-object and articulated HOI from surface keypoint dynamics","Contact distance fields guide whole-body motion from keypoint flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3183,"prompt_tokens":957,"completion_tokens":2226,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":573,"tokens_out":2226,"duration_ms":15234,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:51:53.970736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prompt a trained model to open a prismatic drawer in a cabinet it has not seen, fit a line to each generated drawer trajectory, and measure the per-frame deviation of the drawer's recovered pose from that line and from the cabinet's track. If the median deviation exceeds the paper's own 2 cm assembly-connectivity threshold, or if the recovered joint value leaves the ground-truth range by more than 10% for a substantial fraction of frames, the claim that joint mechanisms are learned from keypoint trajectories alone is refuted; the paper's own appendix reports a 5.20% frame-level connectivity violation rate, so the test should show that rate stays near zero across novel articulated objects.","supporting_citations":[{"cited_title":"2025 , booktitle =","cited_arxiv_id":null,"evidence_quote":"It is the closest prior contact distance field; the paper extends its single-object, skeleton-based field to whole-body multi-object settings and compares against it."}],"review_version":2}