{"id":"f8005e62-5737-470e-919b-9907b7d037a9","arxiv_id":"2606.18243","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"MOCHI enhances noisy collaborative human-object interaction captures via grasp optimization followed by diffusion-based full-body refinement that incorporates interaction information into single-person motion priors.","lead":"MOCHI is a two-stage system that first optimizes realistic hand grasps from noisy body poses in multi-person object interactions and then refines full-body motions using a diffusion model guided by single-person priors plus interaction terms. A smart generalist might read it to see how post-processing can turn lower-quality motion capture into usable data for robotics, animation, or VR involving groups of people handling objects together.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Single-person motion priors + interaction objectives may fail to encode mutual anticipation in MHOI","rationale":"The reader’s weakest assumption directly isolates the methodological hinge of the pipeline. The abstract and claim language provide no additional evidence that the objectives close the multi-agent gap, so the load-bearing risk remains unchanged even with full text available.","tokens_in":1800,"tokens_out":322,"duration_ms":20530,"concrete_test":"Select 5–10 MHOI sequences with explicit mutual anticipation (e.g., one participant reaching in response to the other’s preparatory motion). After running the full MOCHI pipeline, measure (a) cross-participant velocity correlation at contact events and (b) temporal offset between preparatory and responsive actions; compare against the original noisy captures and any available multi-person ground truth. A statistically significant drop in these metrics relative to single-person baselines would confirm the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the diffusion-based refinement stage (using single-person priors augmented by human-object and human-human objectives) produces motions that preserve collaborative dynamics. The priors originate from individual motion data and therefore lack built-in modeling of inter-participant anticipation or continuous mutual adjustment. The added objectives are described only at a high level (contact consistency, semantic alignment) and may enforce only local constraints rather than the global coordination needed for collaborative sequences. If this gap exists, the pipeline can remove capture artifacts while still introducing or preserving dynamics-breaking artifacts, undermining the effectiveness and robustness claims across participant counts and interaction types.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces MOCHI, a two-stage framework for enhancing noisy collaborative multi-human object interaction (MHOI) captures. Stage 1 optimizes physically plausible and semantically consistent hand grasps from noisy body inputs and extends them into full hand-object sequences. Stage 2 refines full-body motions for all participants via diffusion-based noise optimization that augments single-person motion priors with additional objectives encoding human-object and human-human interaction information. The authors claim the pipeline is effective on data from existing capture methods or generative models, robust across varying participant counts and interaction types, and enables applications such as keyframe-based MHOI creation and data augmentation via object geometry variation.","tokens_in":1919,"tokens_out":601,"duration_ms":34477,"significance":"If the central claims hold with rigorous quantitative support, the work would be significant for computer vision and graphics by providing a practical method to improve the quality of MHOI datasets, a known bottleneck for modeling complex collaborative dynamics. The combination of grasp optimization with diffusion priors augmented by interaction objectives represents a targeted approach to artifact removal while attempting to preserve multi-agent coordination; successful validation could directly benefit downstream tasks in animation, robotics, and interaction synthesis.","major_comments":[{"comment":"Abstract: the central claims of effectiveness and robustness across participant numbers and interaction types are asserted without any quantitative metrics, baseline comparisons, error bars, or ablation results; this absence makes it impossible to assess whether the pipeline actually preserves collaborative dynamics or merely removes local artifacts.","section":"Abstract"},{"comment":"Diffusion-based refinement stage (described in the method): the approach augments single-person motion priors with high-level objectives for contact consistency and semantic alignment, but provides no derivation or validation showing that these objectives encode mutual anticipation and continuous inter-participant adjustment rather than only local constraints; because the priors originate from individual motion data, this gap is load-bearing for the claim that refined motions maintain collaborative dynamics across diverse interaction types.","section":"Method (diffusion-based noise optimization framework)"},{"comment":"Experimental results section: the robustness claim across varying numbers of participants and interaction types requires explicit cross-condition quantitative evaluation (e.g., metrics stratified by participant count or interaction category) with statistical significance; without such breakdowns or comparisons to single-person-only baselines, the added interaction objectives' contribution cannot be isolated.","section":"Experiments"}],"minor_comments":[{"comment":"The description of how optimized grasps are 'extended into complete hand-object interaction sequences' lacks implementation details on temporal consistency enforcement.","section":"Stage 1"},{"comment":"Notation for the interaction objectives (e.g., weights on human-object vs. human-human terms) should be formalized with explicit equations to allow reproducibility.","section":"Method"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive feedback. We will revise the manuscript to strengthen the presentation of quantitative results and provide additional clarifications in the method and experiments sections.","responses":[{"response":"The abstract summarizes the paper's claims, while detailed quantitative support, including metrics, baselines, and ablations, is provided in the experiments section. We agree this could be better highlighted and will update the abstract to include key quantitative findings on effectiveness and robustness.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claims of effectiveness and robustness across participant numbers and interaction types are asserted without any quantitative metrics, baseline comparisons, error bars, or ablation results; this absence makes it impossible to assess whether the pipeline actually preserves collaborative dynamics or merely removes local artifacts."},{"response":"The interaction objectives are specifically designed to couple the motions of multiple participants through shared contact and semantic terms, thereby encoding collaborative dynamics beyond local constraints. We will add further explanation and examples in the revised method section to validate how these objectives promote inter-participant adjustment.","revision_made":"partial","referee_comment":"[Method (diffusion-based noise optimization framework)] Diffusion-based refinement stage (described in the method): the approach augments single-person motion priors with high-level objectives for contact consistency and semantic alignment, but provides no derivation or validation showing that these objectives encode mutual anticipation and continuous inter-participant adjustment rather than only local constraints; because the priors originate from individual motion data, this gap is load-bearing for the claim that refined motions maintain collaborative dynamics across diverse interaction types."},{"response":"We have evaluated on data with varying participant counts and interaction types, with overall results supporting robustness. To address the request for stratified evaluation, we will include additional breakdowns by participant number and interaction type in the experiments section, along with comparisons to single-person baselines and statistical analysis.","revision_made":"yes","referee_comment":"[Experiments] Experimental results section: the robustness claim across varying numbers of participants and interaction types requires explicit cross-condition quantitative evaluation (e.g., metrics stratified by participant count or interaction category) with statistical significance; without such breakdowns or comparisons to single-person-only baselines, the added interaction objectives' contribution cannot be isolated."}],"tokens_in":1539,"tokens_out":444,"duration_ms":41299,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"MOCHI is a two-stage method that first optimizes hand grasps from noisy body input to make them physically plausible and semantically consistent, then refines the full-body motions of all participants with a diffusion model that starts from single-person priors and adds objectives for human-object contact and human-human interaction.\n\nThe new piece is the specific combination for MHOI: grasp optimization followed by interaction-augmented diffusion to handle simultaneous human-human and human-object issues in one pipeline. It does a clear job naming the common artifacts (contact misalignment, jitter, incomplete fingers) and sketching how each stage targets them. The approach of extending single-person motion models rather than building everything from multi-person data from scratch is a reasonable engineering choice for this area.\n\nThe soft spot is the complete absence of quantitative evidence. The abstract states that the pipeline is effective and robust across capture methods, generative data, participant counts, and interaction types, yet it shows no metrics, no baselines, no ablations, and no details on how the interaction objectives were chosen or validated. Without those, it is impossible to judge whether the added terms actually preserve collaborative dynamics or merely enforce local constraints.\n\nThe stress-test point about single-person priors struggling with mutual anticipation lands as a real open question here, since the objectives are described only at a high level.\n\nThis paper is for researchers who need practical tools to clean or augment MHOI datasets for animation, robotics, or generative modeling. A reader already working on motion priors or interaction capture would find the pipeline description useful if the experiments hold up.\n\nI would send it for peer review so the full results and implementation details can be checked.","headline":"MOCHI gives a concrete two-stage pipeline for fixing MHOI capture artifacts but the abstract supplies no numbers or baselines to back its effectiveness claims.","tokens_in":2424,"tokens_out":410,"would_cite":false,"duration_ms":32667,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"MOCHI enhances noisy collaborative human-object interaction data through grasp optimization and diffusion-based motion refinement using single-person priors.","keywords":["motion enhancement","human-object interaction","collaborative interactions","diffusion models","motion optimization","grasp generation","MHOI data"],"falsifier":"Observing whether the optimized motions maintain consistent contacts and smooth trajectories in long sequences of complex multi-person object manipulations that were not used in training.","tokens_in":2678,"feed_emoji":"🤝","tokens_out":489,"duration_ms":29325,"temperature":0.7,"pith_summary":"The paper introduces MOCHI, a two-stage framework designed to clean up artifacts in MHOI captures such as hand-object misalignments and motion jitter. In the first stage, it optimizes physically plausible hand grasps from noisy body poses and extends them to full sequences. The second stage refines the full-body motions of all participants via a diffusion model that incorporates interaction constraints into single-person motion priors. This matters because high-quality motion data is essential for training models that simulate realistic collaborative scenarios involving multiple people and shared objects. The method demonstrates robustness to different data sources and participant numbers while enabling new applications like keyframe editing.","feed_headline":"Two-stage optimization cleans noisy multi-human object motions","feed_subtitle":"Grasp optimization followed by diffusion refinement with single-person priors produces plausible collaborative sequences from noisy inputs.","key_machinery":"Diffusion-based noise optimization framework that encodes human-object and human-human interactions into single-person motion priors.","core_discovery":"MOCHI is a two-stage pipeline that first generates physically plausible hand grasps through optimization from noisy body input and extends them into complete hand-object sequences, then refines full-body motions for all participants using a diffusion-based noise optimization framework augmented with objectives that encode human-object and human-human interaction information within single-person motion priors.","pith_inferences":["If single-person priors can be augmented this way, similar techniques might apply to other multi-agent motion problems like team sports.","The approach could reduce reliance on specialized multi-person capture equipment.","Extending the method to real-time applications might improve interactive simulations in VR."],"forward_implications":["Works on data from existing capture methods or generative models.","Robust across varying numbers of participants and interaction types.","Supports applications such as keyframe-based MHOI creation.","Enables data augmentation by varying object geometries."],"fun_headline_variants":["Optimization pipeline fixes noisy multi-human object interactions","Hand grasp optimization precedes diffusion-based motion refinement","Single-person priors enable MHOI motion enhancement via diffusion","Two-stage MOCHI framework cleans collaborative human-object captures"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Single-person motion priors can be augmented with additional objectives to encode the mutual anticipation and adjustments in collaborative interactions without introducing breaking artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Optimization pipeline fixes noisy multi-human object interactions","Hand grasp optimization precedes diffusion-based motion refinement","Single-person priors enable MHOI motion enhancement via diffusion","Two-stage MOCHI framework cleans collaborative human-object captures"]},"model":"grok-4.3","cost_usd":0.007222,"raw_usage":{"total_tokens":3349,"prompt_tokens":705,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":72224500,"prompt_tokens_details":{"text_tokens":705,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2586,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":705,"tokens_out":58,"duration_ms":27506,"temperature":1.0,"reasoning_tokens":2586,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T01:19:07.872081+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Observing whether the optimized motions maintain consistent contacts and smooth trajectories in long sequences of complex multi-person object manipulations that were not used in training.","supporting_citations":[],"review_version":1}