{"id":"25f0cd3d-d899-44b1-aaa7-e865fbfd8c24","arxiv_id":"2505.21437","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CoDA generates coordinated whole-body articulated-object manipulation by optimizing the noise of three decoupled diffusion models, guided by BPS-based end-effector and object trajectories.","lead":"This preprint introduces CoDA, a pipeline that generates full body motion, both hands, and articulated object motion from text or target poses by optimizing the noise of three separate diffusion models. It claims to be the first method to jointly synthesize body, hand, and articulated object motion, with applications in virtual humans and robotics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoupled-marginal assumption in §4.3 is untested: no objective or metric scores joint plausibility of the assembled body+hand pose, so 'coordination emerges' is not yet demonstrated.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: three marginal diffusion priors are combined through gradient-based noise optimization, and the paper never directly checks that the resulting joint distribution is the whole-body manipulation distribution. I sharpen this into a concrete, testable interface-consistency problem: the optimization loss in Eq. (9) contains no term that scores the joint plausibility of the assembled body+hand pose, and the body and hand priors share no training signal at the wrist. The empirical results are plausible and the ablation study provides indirect support for the decoupled design, but the central mechanism remains unverified. The lack of error bars, code, and a larger user study are secondary confirmability issues; they do not by themselves threaten the core claim. Because the concern is an untested assumption rather than a demonstrated failure, keeping the verdict CONDITIONAL is appropriate; no verdict adjustment is needed.","tokens_in":21584,"tokens_out":7168,"duration_ms":93109,"concrete_test":"On held-out ARCTIC/GRAB sequences, train an unconditional whole-body diffusion model (or use a pretrained whole-body motion prior) on the same training data, including body and hands together. For each CoDA-generated, baseline-generated, and real sequence, compute the average per-frame denoising error (or negative log-likelihood) under this reference model, and separately compute the wrist-orientation error between the body model's wrist frame and the wrist frame implied by the optimized hand configuration. If CoDA samples fall inside the reference model's real-data error distribution and wrist-orientation errors match real motion, the decoupled prior is validated; if they are outliers, the joint-prior assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is §4.3: three independently trained diffusion models (body on AMASS, hands on ARCTIC/GRAB) are combined by minimizing L in Eq. (9) over z={zb,zlh,zrh}. This implicitly assumes that optimizing in the product of marginal noise spaces, with only end-effector, penetration, and foot losses, lands on the joint whole-body manipulation manifold. No term in Lee (Eq. 10), Lpen (Eq. 11), or Lreg (Eq. 12) evaluates the joint plausibility of the assembled pose; the diffusion priors regularize each component separately. Because the body prior and hand priors share no training signal at the wrist interface, the optimizer can satisfy the targets with a body pose that is individually likely and a hand pose that is individually likely while their combination is jointly unlikely (e.g., a wrist/palm orientation inconsistent with the finger contacts, or an unnatural shoulder compensation). The reported FID and the 16-participant user study are aggregate and do not isolate interface consistency. Ablation (e) shows that replacing the decoupled models with a single whole-body model hurts FID (2.699 vs 2.283) and physical metrics, which supports the decoupled design but does not test whether the product-of-marginals prior is faithful. Thus the central claim that coordination 'naturally emerges' is plausible but not directly verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoDA, a three-stage framework for synthesizing whole-body manipulation of articulated objects. Given an initial human/object pose and text, it first generates an articulated object trajectory, then predicts end-effector trajectories using a unified basis-point-set (BPS) representation, and finally produces whole-body motion by optimizing the input noise of three independently trained diffusion models for the body, left hand, and right hand. The optimization losses track end-effector positions, penalize hand-object penetration, and regularize foot contact. Experiments on ARCTIC and GRAB report state-of-the-art FID, physical-plausibility, and user-study results, with ablations and demonstrations of object-pose control, walking while manipulating, and whole-body generation from hand-only data.","tokens_in":21919,"tokens_out":4704,"duration_ms":51814,"significance":"If validated, the paper makes a significant contribution: it is the first to jointly generate body, hand, and articulated-object motion for whole-body manipulation, and the decoupled diffusion design is a practical way to leverage large body-only and hand-only datasets. The BPS-based end-effector representation is a clean idea that ties trajectory prediction to object geometry, and the noise-space optimization over multiple models is a meaningful extension of DNO. However, the central empirical claims currently rest on aggregate metrics without error bars, an under-specified baseline extension protocol, a small user study, and no direct test of the joint plausibility of the assembled body-plus-hand poses.","major_comments":[{"comment":"The load-bearing assumption—that optimizing over the product of three independently trained marginal diffusion priors, using only the end-effector, penetration, and regularization losses, lands on the joint whole-body manipulation manifold—is not directly tested. None of the terms in Eq. (10), Eq. (11), or Eq. (12) evaluates the joint plausibility of the assembled body-plus-hand pose, and the body and hand priors share no training signal at the wrist interface. Ablation (e) shows that a single whole-body model performs worse (FID 2.699 vs 2.283, IV 12.45 vs 5.25), but this compares two architectures under the same optimization and does not verify that the product-of-marginals prior is faithful. I request a direct interface-consistency evaluation (e.g., wrist/palm orientation consistency with finger contacts, or distance of assembled poses to the real whole-body manifold) so that the 'coordination naturally emerges' claim is supported by a metric rather than by aggregate FID and a 16-participant study.","section":"§4.3 (Eqs. 7-12) and §5.4 (Table 4)"},{"comment":"The state-of-the-art comparison is not fully reproducible because the protocol for extending IMoS, MDM, OMOMO, Text2HOI, and CHOIS to the whole-body articulated-object task is not described. It is not specified whether the same object-motion and end-effector generation models are used for all baselines, how the hand components are added, or which training data are used. In addition, no error bars, standard deviations, or significance tests are reported for any metric; since the sample sizes are not stated, the reader cannot assess the stability of the reported differences (e.g., FID 2.283 vs 3.710 for OMOMO on ARCTIC). Please report variances over multiple seeds or bootstrap intervals and describe the extension protocol in detail.","section":"§5.3 and Tables 1 and 3"},{"comment":"The user study has 16 participants but no protocol details are given: the number of sequences evaluated, the number of pairwise comparisons per participant, the presentation order, the expertise of the participants, and the statistical test used are all missing. Since the claim that the method 'significantly outperforms' in perceptual quality rests on this table, please report the full protocol and a significance measure (e.g., a p-value or a confidence interval for the preference rates).","section":"Table 2 and §5.2"},{"comment":"The optimization schedule is hand-tuned with staged thresholds (300/500/800 steps) and loss weights λee=1, λpen=5.0, λreg=1.0, yet no sensitivity analysis is provided. The reported improvements in Table 4 depend on these choices; a small study varying the schedule and weights would establish that the method is not overfit to these specific settings.","section":"§D.2"}],"minor_comments":[{"comment":"The notation ∥min(Jy) − 0.02∥1 is ambiguous; please specify whether the minimum is over all joints, over each frame, or over the foot joints, and define the norm explicitly.","section":"Eq. (12)"},{"comment":"There are minor punctuation errors, e.g., 'DiffH2O [13] applies ... interactions..' contains a double period.","section":"§2"},{"comment":"The R-Precision metric is not defined precisely; please state the retrieval procedure and the threshold used for Top-3 accuracy.","section":"§5.2"},{"comment":"The text says 'As shown in figure 3' but the figure is referenced elsewhere as 'Figure 3'; please use consistent capitalization.","section":"§4.2"},{"comment":"The phrase 'All diffusion models use 1000 sampling steps during training' conflates the number of diffusion steps in training with inference-time sampling; please rephrase to refer to the forward diffusion noise schedule.","section":"§D.1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core idea is interesting, but the experimental validation needs strengthening before acceptance. In particular, I would ask the authors to add a direct joint-plausibility test for the assembled body-hand output, report error bars and full baseline-extension details, and substantially document the user study. The 'first work' claim is defensible if the descriptions of TOHO and IMoS are accurate, but it should be softened if any of those methods can handle articulated objects."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Pi, Cen, Dou, and Komura have written the first paper I know of that generates whole-body manipulation of articulated objects—body, hands, and object motion together. The method is a coordinated noise-space optimization over three decoupled diffusion models (body, left hand, right hand), fed by object and end-effector trajectories generated by two conditional diffusion models. The decoupled training is the clever part: body model sees AMASS plus the interaction datasets, hand models see ARCTIC/GRAB, so each component gets more data than any single joint model would. The BPS-based end-effector representation is a reasonable way to tie hand targets to object geometry, and the RoPE object encoding follows CaPE cleanly. The ablations are informative, and the limitations section is honest—16.93 minutes per sequence is slow, and they say so.\n\nThe central claim is plausible: coordination \"naturally emerges\" from gradient flow through the kinematic chain. But the paper does not directly test the assumption that the product of the three marginal priors lands on the joint whole-body manifold. There is no metric for wrist-hand interface consistency, so a body pose that is individually likely and a hand pose that is individually likely could combine into an unlikely whole. The relative fingertip-to-wrist term in Lee helps, and the hand models were trained on whole-body captures, so the concern is partially mitigated, but it remains untested. The ablation replacing the decoupled models with a single whole-body model (e) shows the decoupled version has better FID, which supports the design, but that does not test interface plausibility.\n\nThe bigger experimental problems are simpler. No error bars anywhere. Tables 1, 3, and 4 report single runs, and FID differences of 0.2–0.4 could be noise. The baselines are extended without a described protocol; I cannot tell how IMoS or MDM were adapted to produce hands and object motion. The user study has 16 participants. Code is promised but not released. None of these are fatal, but they make the SOTA claim thinner than the prose implies.\n\nWho this is for: anyone working on human-object interaction generation, especially articulated objects. It deserves a serious referee. I would send it out, but ask for error bars, baseline extension details, and one direct evaluation of body-hand interface consistency—say, a per-sequence check of wrist orientation against finger contact or a user study isolating interface errors.","headline":"First whole-body articulated-object manipulation generator with a plausible decoupled-diffusion recipe, but the coordination claim is not directly measured and the experimental evidence lacks error bars.","tokens_in":22391,"tokens_out":2761,"would_cite":true,"duration_ms":31075,"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":"This paper proposes coordinated diffusion noise optimization of three decoupled diffusion models to generate whole-body manipulation of articulated objects, with better motion quality and physical plausibility than adapted baselines on…","keywords":["whole-body motion generation","hand-object interaction","articulated object manipulation","diffusion models","noise-space optimization","basis point sets","SMPL-X"],"falsifier":"A decisive test is to train the body model only on locomotion data and the hand models only on grasping data from a different object set, then run the full pipeline; if end-effector tracking accuracy and contact ratio remain high, the optimization alone creates coordination, and if they collapse, the coordinated manifold must already exist in the training data.","tokens_in":21426,"feed_emoji":"🖐️","tokens_out":8270,"duration_ms":81591,"temperature":0.7,"pith_summary":"The paper is trying to establish that whole-body manipulation of articulated objects—body, both hands, and the moving object—can be generated as a single coordinated sequence by optimizing the input noise of three independently trained diffusion models, one for the body, one for the left hand, and one for the right hand. Coordination emerges from gradient flow along the human kinematic chain: hand objectives pull the torso and limbs into the right posture, while body objectives keep the hands consistent with the object. The claim matters because it decouples data requirements: the body model is trained on large general motion corpora and the hand models on hand-specific manipulation datasets, so no large paired whole-body manipulation corpus is needed to generate such interactions. The paper reports that the framework, called CoDA, outperforms adapted baselines on motion quality, foot skating, and hand-object contact metrics on both the ARCTIC and GRAB benchmarks, and it enables object pose control, walking while manipulating, and whole-body generation from hand-only trajectories.","feed_headline":"Optimizing diffusion noise coordinates body, hands, and object motion","feed_subtitle":"Body and hand priors trained separately are fused at inference, unlocking new manipulation skills.","key_machinery":"The load-bearing mechanism is coordinated diffusion noise optimization over decoupled motion priors. Three diffusion models—body $M_b$, left hand $M_{lh}$, right hand $M_{rh}$—are trained on separate marginal datasets, and inference solves $z^* = \\arg\\min_z L(f(z))$, where $f(z)$ maps the concatenated noise through the diffusion models and forward kinematics to global joint positions; the loss $L = \\lambda_{ee} L_{ee} + \\lambda_{pen} L_{pen} + \\lambda_{reg} L_{reg}$ enforces end-effector tracking, penetration avoidance, and foot contact. The second supporting object is the shared basis point set (BPS) representation, which encodes both object geometry and the 12 end-effectors (wrists and fingertips) as distances to the same basis points, so the trajectory targets live in the same reference frame as the object. Coordinated body adaptation is attributed to gradients flowing from hand objectives through the kinematic chain during optimization.","core_discovery":"CoDA's central claim is that joint body-hand-object motion for articulated manipulation can be synthesized by first generating an object trajectory and end-effector trajectories, then solving for the diffusion noise vectors of three decoupled models so that the denoised poses track those trajectories. The end-effector stage uses a unified basis point set (BPS) representation: fingertip and wrist positions are encoded as distances to the same basis points that encode the object geometry, which puts the interaction targets in the object's own reference frame and improves precision. The whole-body stage optimizes $z=\\{z_b, z_{lh}, z_{rh}\\}$ through a loss combining end-effector tracking, hand-object penetration, and foot-contact regularization, with DDIM sampling used so gradients can flow back through the solver. On ARCTIC, CoDA reports FID 2.283 versus 3.710 for the best adapted baseline, interpenetration volume 5.25, and contact ratio 0.086, close to the real-motion value of 0.085; user-study participants preferred it for realism and physical plausibility at rates around 88 percent. The authors state this is the first work to jointly generate body, hand, and articulated object motion for whole-body manipulation.","pith_inferences":["Because coordination is produced by optimization through the kinematic chain rather than by joint training, the same scheme could plausibly extend beyond three modules, for example adding a separate object-motion diffusion model or a face and expression module, provided the objectives remain differentiable.","A testable consequence is that the benefit of the decoupled design should grow as the body and hand training distributions diverge; an ablation that varies the amount of paired whole-body data during training would reveal how much coordination the optimization actually buys.","The 16.9-minute optimization per 300-frame sequence suggests the practical path to real-time use is distilling the optimized-noise behavior into a feed-forward conditional model, with the current pipeline serving as the teacher.","Because the end-effector targets are distances to a shared basis point set, the trajectory stage may transfer across object geometries and categories more easily than the motion stage; evaluating the same trajectory models on unseen object categories would test this directly."],"forward_implications":["If the method holds up, whole-body manipulation can be synthesized without a large paired dataset: the body prior comes from general motion capture and the hand priors from hand-only interaction data, so novel combinations of body and finger motion become possible.","Users can control the object's pose at keyframes, shift the root target to make the character walk while manipulating, and feed wrist and fingertip trajectories from hand-only recordings to obtain full-body motion.","The generated sequences are accurate enough that the framework can output reference motion for physics-based humanoids, which then physically manipulate the object in simulation.","On the ARCTIC and GRAB benchmarks, the approach improves FID, foot skating, interpenetration volume and depth, and contact ratio over the adapted baselines considered.","The authors state this is the first method to jointly generate body, hand, and articulated object motion for whole-body manipulation tasks."],"supporting_citations":[{"why":"Supplies the diffusion-noise optimization methodology that the whole-body stage is built on.","marker":"[40]"},{"why":"Provides the BPS-based object representation and the ARCTIC evaluation protocol the paper follows.","marker":"[132]"},{"why":"Introduces basis point sets, the distance-based geometry representation used for both object and end-effector encoding.","marker":"[81]"},{"why":"AMASS is the large-scale body-only motion dataset used to train the body diffusion model.","marker":"[63]"},{"why":"ARCTIC supplies articulated-object manipulation data for training and is the primary evaluation benchmark.","marker":"[19]"},{"why":"GRAB supplies whole-body grasping data for hand-model training and evaluation on rigid objects.","marker":"[93]"},{"why":"MDM provides the transformer-based diffusion architecture used for all component models.","marker":"[97]"},{"why":"CaPE's relative-pose RoPE encoding is adapted to represent object trajectories in the trajectory diffusion models.","marker":"[44]"}],"fun_headline_variants":["Diffusion noise optimization coordinates whole-body articulated manipulation","CoDA unifies body, hand, and object motion via noise optimization","First joint generation of whole-body and object motion via diffusion","Noise-space optimization fuses body and hand priors for object manipulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that separately trained body and hand priors, when pushed together by noise-space optimization, can land on plausible joint whole-body manipulation rather than on merely individually plausible parts.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion noise optimization coordinates whole-body articulated manipulation","CoDA unifies body, hand, and object motion via noise optimization","First joint generation of whole-body and object motion via diffusion","Noise-space optimization fuses body and hand priors for object manipulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3703,"prompt_tokens":1071,"completion_tokens":2632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":2570}},"tokens_in":687,"tokens_out":2632,"duration_ms":19873,"temperature":1.0,"reasoning_tokens":2570,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:28:01.748487+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to train the body model only on locomotion data and the hand models only on grasping data from a different object set, then run the full pipeline; if end-effector tracking accuracy and contact ratio remain high, the optimization alone creates coordination, and if they collapse, the coordinated manifold must already exist in the training data.","supporting_citations":[{"cited_title":"Bimart: A unified approach for the synthesis of 3d bimanual interaction with articulated objects","cited_arxiv_id":null,"evidence_quote":"Provides the BPS-based object representation and the ARCTIC evaluation protocol the paper follows."},{"cited_title":"Grab: A dataset of whole-body human grasping of objects","cited_arxiv_id":null,"evidence_quote":"GRAB supplies whole-body grasping data for hand-model training and evaluation on rigid objects."},{"cited_title":"Human motion diffusion model","cited_arxiv_id":null,"evidence_quote":"MDM provides the transformer-based diffusion architecture used for all component models."}],"review_version":1}