{"id":"ce17bb67-1325-4f40-aef0-b05890d53acc","arxiv_id":"2509.10096","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new motion capture dataset of physical caregiving assists and an interaction-aware diffusion model that beats baselines at predicting both agents' poses.","lead":"This paper introduces HHI-Assist, a motion capture dataset of caregivers helping care receivers with transfers, plus a diffusion model that predicts both people's future poses. It matters because assistive robots need to anticipate human movement during physical contact, and no marker-based dataset existed for these tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significantly outperforms' and 'strong generalization' claims lack statistical support: Table II reports single-run MPJPE with no variance, and the Task 3 generalization evaluation omits all baseline comparisons.","rationale":"The reader's stated weakest assumption is transferability from human-human to human-robot interaction, but the authors explicitly acknowledge that transfer is future work, so it is not a hidden flaw in the paper's internal argument. The more load-bearing issue is that the central comparative claims lack statistical substantiation: Table II has no error bars or significance tests, and the generalization experiment reports no baseline comparisons. These omissions are addressable and do not negate the dataset contribution, which is a genuinely new marker-based motion capture resource for physical assistance scenarios. The dataset size, participant split, and public availability are positive points. A conditional acceptance requiring a significance analysis and a baseline comparison on Task 3 would settle the concern; thus the reader's CONDITIONAL verdict remains appropriate.","tokens_in":12079,"tokens_out":4610,"duration_ms":41314,"concrete_test":"Re-run the Table II evaluation with IDD and at least the two strongest baselines (TCD and siMLPe) using 10 random seeds, and also evaluate all baselines on Task 3; report mean +/- std and paired bootstrap 95% confidence intervals or Wilcoxon signed-rank p-values over the 8.7k test sequences. If IDD's CI overlaps TCD's on either agent, or if any baseline matches IDD on Task 3, the claims of 'significant' improvement and 'strong generalization' should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section V-D) is that IDD 'significantly outperforms' interaction-unaware baselines, and the abstract claims 'strong generalization to unseen scenarios.' Both halves rest on incomplete evidence. Table II reports one MPJPE value per model/horizon with no standard deviation, confidence interval, number of seeds, or significance test. The largest average improvement over TCD is about 3 mm for the caregiver (50.4 vs 52.0) and 3.2 mm for the care receiver (34.3 vs 37.5), so the word 'significant' is doing real work; with 8.7k test sequences a paired test might well reach p<0.05, but the paper does not show it. For generalization, Section V-E1 reports only IDD absolute errors on Task 3 (89.3/62.5 mm average for CG/CR), with no comparison to Zero-Vel, Constant-Vel, or the learned baselines, so there is no evidence that IDD generalizes better than a trivial predictor. Task 3 is also only 10 demonstrations from a single pair, making the conclusion fragile. These are omissions rather than contradictions, but they directly undercut the strongest sentence of the paper. The transferability-to-HRI premise, by contrast, is explicitly flagged as future work in the Introduction and is not required for the dataset's internal validity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HHI-Assist, a motion-capture dataset of human-human physical assistance interactions (sit-to-stand, lay-to-sit, lay-to-stand, and unconstrained tasks), and proposes IDD, a conditional Transformer-based denoising diffusion model that predicts future poses of both the caregiver and care receiver conditioned on the observed poses of both agents. The authors evaluate IDD against several baselines on Tasks 1-2, report a generalization experiment on the small Task 3, and study joint-angle versus joint-position representations. The abstract claims improvements over baselines and strong generalization to unseen scenarios, and the dataset and code are publicly released.","tokens_in":12368,"tokens_out":3252,"duration_ms":28822,"significance":"If the claims are supported, HHI-Assist would be a useful new resource: it is, to my knowledge, the first marker-based motion-capture dataset focused specifically on physical assistance scenarios, with a reasonable number of demonstrations from multiple participant pairs. The IDD model is a natural extension of the authors' earlier single-agent diffusion pose predictor, and the dataset-plus-benchmark package with released code could be of value to researchers in physical human-robot interaction, behavior cloning, and human motion forecasting. The paper also contains useful exploratory analyses, including t-SNE task separability and link-length consistency for joint-angle representations. However, the headline claims are currently stronger than the evidence: the main table has no variance or significance testing, and the generalization experiment lacks baseline comparisons. These issues are fixable but are load-bearing for the paper's central message.","major_comments":[{"comment":"The central claim that IDD 'significantly outperforms' the interaction-unaware baselines is not supported by the reported statistics. Table II gives a single MPJPE value per model and horizon with no error bars, confidence intervals, number of seeds, or significance tests. The average margins over TCD are about 1.6 mm for the caregiver and 3.2 mm for the care receiver; with a test set of 8.7k sequences a paired test might well show significance, but the paper does not demonstrate this. Please report variances across seeds or bootstrapped confidence intervals and perform significance tests (or at least state whether the differences are significant), or soften the wording from 'significantly outperforms' to 'achieves lower average MPJPE.'","section":"Section V-D, Table II"},{"comment":"The generalization evaluation does not support the abstract's claim of 'strong generalization to unseen scenarios.' Only IDD absolute errors on Task 3 are reported (average MPJPE 89.3 mm for CG and 62.5 mm for CR), with no comparison to Zero-Vel, Constant-Vel, or any learned baseline. Under distribution shift, trivial predictors can be competitive, so the reader cannot tell whether IDD generalizes better than a constant-velocity model. In addition, Task 3 consists of only 10 demonstrations from a single participant pair, making the conclusion fragile. Please either add baseline comparisons on Task 3 and frame the result as preliminary, or substantially scale up the generalization evaluation.","section":"Section V-E1"},{"comment":"The adaptation of the single-agent baselines to the two-agent interaction setting is underspecified. In particular, TCD is described as a denoising diffusion model for single human pose prediction, and DSTFormer as a dual-stream spatio-temporal Transformer, but the paper does not state how these models receive the concatenated CG/CR observations or how their output heads predict both agents. Without this information the benchmark comparisons cannot be reproduced exactly. Please specify the input/output encoding and any hyperparameter changes for each baseline, including siMLPe.","section":"Section V-B and V-C"},{"comment":"The procedure for constructing the 44.8k/3.5k/8.7k train/validation/test sequences from the 908 demonstrations should be described. If sliding windows are used, the stride and the relationship between test sequences and training sequences from the same demonstration clip need to be stated; the no-participant-overlap guarantee does not by itself rule out clip-level overlap between temporally adjacent training and test windows. This is important for the validity of the benchmark numbers in Table II.","section":"Section V-A"}],"minor_comments":[{"comment":"In the notation, 'GR' appears instead of 'CR' as the abbreviation for care receiver.","section":"Section IV-A"},{"comment":"There is a doubled comma in 'MPJPE, , likely'; also, 'around 7% higher average MPJPE' should be clarified as a relative rather than absolute increase.","section":"Section V-E3"},{"comment":"The baseline referred to as DSTFormer is reference [60], whose title is 'MotionBERT: A Unified Perspective on Learning Human Motion Representations'; the naming should be reconciled in the text and Table II.","section":"References"},{"comment":"The phrase 'a non-monetary incentive equivalent to 10 USD/h' is awkward; 'USD 10/hour' would be clearer.","section":"Section III-A"},{"comment":"In the Delayed IDD experiment, the time indices 't=0 to t=1' and 't=0.5 to t=1.5' should be explicitly labeled as seconds (or frames) to avoid ambiguity.","section":"Section V-E2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an accepted RA-L preprint, and the dataset release is a genuinely useful contribution. The main revision need is statistical support for the comparative claims and baseline comparisons in the generalization experiment; both are well within the scope of a revision. The transferability-to-HRI premise is explicitly labeled as future work and should not be treated as a standalone reason to reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the dataset is the real contribution and it's a good one; the model is a reasonable but incremental baseline, and the paper oversells 'significant' and 'strong generalization' without the statistical backing. I'd send it to review but ask for revisions.\n\nThe HHI-Assist dataset fills a genuine gap. Existing close-contact datasets (Hi4D, CHI3D, AIR-Act2Act) don't target physical assistance, and none use marker-based mocap with this kind of caregiver-care receiver transfer activity. 908 clips across four tasks, with 11-13 pairs for the two core tasks, a held-out test set with no participant overlap, and clear documentation of the collection protocol (Optitrack, 120 Hz, BVH). That's a solid, reusable resource for assistive robotics and motion prediction.\n\nThe IDD model is a direct extension of the authors' own TCD model to two agents, conditioning on both partners' histories. That's disclosed and appropriate. It beats the baselines on the test set by a small but consistent margin (e.g., average MPJPE 50.4 vs 52.0 for CG, 34.3 vs 37.5 for CR). The word 'significantly' in Section V-D is doing work the paper doesn't back up: Table II has single-run numbers with no variance, no seeds, no significance test. With 8.7k test sequences a paired test would likely be significant, but they need to show it. The generalization story is weaker: Task 3 has only 10 demonstrations from a single pair, and Section V-E1 reports only IDD's absolute errors with no baseline comparison, so 'strong generalization' is unsupported. It's an omission, not a contradiction, but it undercuts the abstract's strongest sentence.\n\nThe transferability-to-HRI premise is explicitly flagged as future work in the Introduction, so I don't hold that against the paper; the dataset stands on its own as a benchmark for human-human assistance, and it's reasonable to posit downstream use in pHRI. The citation pattern is fine. Minor issues: the joint-angle comparison is a nice sanity check on link lengths, though the position model wins on MPJPE as expected. The delayed-coupling experiment is interesting but tangential; it could be cut or better motivated.\n\nBottom line: the dataset alone justifies peer review. The model claims need either softer language or proper error bars and baselines in the generalization experiment. I'd recommend conditional acceptance with those fixes.","headline":"New marker-based HHI assistance dataset is the real contribution; the IDD model is an incremental baseline and the 'significant/strong generalization' claims need statistical support.","tokens_in":12905,"tokens_out":2081,"would_cite":true,"duration_ms":17133,"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":"The paper presents HHI-Assist, a marker-based motion-capture dataset of people physically assisting each other in care tasks, and an interaction-aware diffusion model that predicts both partners' future poses more accurately than…","keywords":["human motion prediction","physical human-robot interaction","assistive robotics","motion capture dataset","denoising diffusion model","interaction-aware pose forecasting","caregiver care receiver","robot learning"],"falsifier":"Recompute MPJPE per test clip and run a paired significance test (e.g., Wilcoxon signed-rank) between IDD and the strongest baseline, siMLPe for the caregiver and TCD for the care receiver; if the reported 1.3 to 3.2 mm average advantages are not significant across the 8.7k test sequences, the claim that the interaction-aware model significantly outperforms its baselines collapses.","tokens_in":11914,"feed_emoji":"🦾","tokens_out":6500,"duration_ms":50142,"temperature":0.7,"pith_summary":"This paper is trying to establish that motion prediction for physical assistance—where one person helps another stand up, sit up, or move—should treat the two people as one coupled system, and that a generative model can learn that coupling from motion-capture data. To that end the authors collected HHI-Assist, 908 recordings of caregiver-care receiver pairs doing sit-to-stand, lay-to-sit, lay-to-stand, and unconstrained movements, with joint positions recorded at 120 Hz. They then built an interaction-aware denoising diffusion model (IDD) that predicts the future poses of both agents conditioned on both of their observed poses. On the benchmark the model reports lower prediction error than single-person baselines for both caregiver and care receiver, and it degrades gracefully on a transfer task it was not trained on. If the results hold, robots assisting people could react to a care receiver's motions earlier and more safely.","feed_headline":"To predict assisted motion, model both bodies, not one","feed_subtitle":"A marker-based dataset of 908 caregiver-care receiver clips backs a diffusion model that beats interaction-unaware baselines.","key_machinery":"The central object is a conditional denoising diffusion model: a Transformer with alternating temporal and spatial multi-head attention blocks, trained to predict the noise that was added to future pose sequences of both agents, conditioned on the concatenated observed pose sequences of both agents. At inference, the network iteratively denoises random Gaussian samples ($T=50$ steps) into pose futures that are consistent with the observed interaction. The supporting machinery is the HHI-Assist dataset itself, which supplies realistic, paired caregiver-care receiver motion for training and benchmarking.","core_discovery":"The central claim is that conditioning pose prediction on both interacting agents' observed motion, rather than on one agent alone, captures the coupled dynamics of physical assistance and yields measurably better pose forecasts. On HHI-Assist, the IDD model achieves average per-joint errors of 50.4 mm for the caregiver and 34.3 mm for the care receiver over a one-second horizon, beating the best interaction-unaware baselines (siMLPe at 51.7 mm and TCD at 37.5 mm, respectively). The model also generalizes to the unseen lay-to-stand task, holding average errors near 89 mm and 63 mm. In a controlled 'delayed' variant where one person's observation is shifted by 0.5 s, errors drop further, supporting the interpretation that the gains come from modeling reciprocal influence.","pith_inferences":["A per-phase error analysis, splitting the horizon into contact onset, lifting, and settling, could reveal whether the interaction-aware advantage concentrates in the moments right after physical contact, which would guide where a robot should pay most attention.","If the transferability assumption holds, the dataset can be retargeted to a robot's kinematics and used directly as training data for behavior cloning of assistance motions, an extension the paper names but does not test.","The rare 'unconstrained' clips (9 demonstrations) could serve as a domain-randomization prior for robustness training, letting downstream models see more varied motion than the four scripted tasks provide.","The measured advantage over the closest baselines is a few millimeters in average MPJPE, so the practical significance for robot controllers is best judged by downstream tasks, such as contact-force prediction, rather than by the raw error gap alone."],"forward_implications":["Controllers for assistive robots can use IDD-style predictions as a look-ahead signal, smoothing reactions to a care receiver's onset of movement during transfers.","The dataset gives future work a fixed benchmark with a train/validation/test split that has no participant overlap, so pose-prediction methods can be compared fairly on physical-assistance motion.","A joint-angle-trained variant produces predictions that keep link lengths constant (unlike joint positions, which drift 7 to 15 mm), a property needed if predictions are converted into robot joint commands.","The model's lower error on the unseen lay-to-stand task suggests the learned coupling transfers partly across tasks, even though the gap between tasks remains a challenge."],"supporting_citations":[{"why":"Base architecture IDD extends: a denoising-diffusion Transformer for single-person pose prediction; IDD adds conditioning on the partner's poses.","marker":"[44]"},{"why":"Supplies the DDPM noising/denoising formulation and loss that the conditional diffusion process is built on.","marker":"[17]"},{"why":"siMLPe, the strongest MLP baseline that IDD must beat for the caregiver (51.7 vs 50.4 mm average MPJPE).","marker":"[16]"},{"why":"DSTFormer/MotionBERT, a spatio-temporal Transformer baseline that IDD outperforms, establishing the contribution over strong sequence models.","marker":"[60]"},{"why":"Hi4D, a close-contact interaction dataset that HHI-Assist contrasts with as the first marker-based motion-capture dataset for physical assistance.","marker":"[56]"}],"fun_headline_variants":["Human-human assist dataset trains diffusion that beats single-agent models","Both bodies, not one: diffusion predicts assistance motion","New benchmark of assistive touch improves robot motion forecasts","Diffusion model learns coupled caregiver-care receiver dynamics","Dataset of human assistance motion boosts pose prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the assumption that how two humans move while helping each other is similar enough to how a human and a robot would move together that predictions learned from human-human data will still help a robot assist a person.","fun_headline_variants_meta":{"raw":{"variants":["Human-human assist dataset trains diffusion that beats single-agent models","Both bodies, not one: diffusion predicts assistance motion","New benchmark of assistive touch improves robot motion forecasts","Diffusion model learns coupled caregiver-care receiver dynamics","Dataset of human assistance motion boosts pose prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1245,"prompt_tokens":880,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":496,"tokens_out":365,"duration_ms":370166,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:57:04.823431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute MPJPE per test clip and run a paired significance test (e.g., Wilcoxon signed-rank) between IDD and the strongest baseline, siMLPe for the caregiver and TCD for the care receiver; if the reported 1.3 to 3.2 mm average advantages are not significant across the 8.7k test sequences, the claim that the interaction-aware model significantly outperforms its baselines collapses.","supporting_citations":[{"cited_title":"A generic diffusion-based approach for 3d human pose prediction in the wild","cited_arxiv_id":null,"evidence_quote":"Base architecture IDD extends: a denoising-diffusion Transformer for single-person pose prediction; IDD adds conditioning on the partner's poses."},{"cited_title":"Motionbert: A unified perspective on learning human motion representations","cited_arxiv_id":null,"evidence_quote":"DSTFormer/MotionBERT, a spatio-temporal Transformer baseline that IDD outperforms, establishing the contribution over strong sequence models."},{"cited_title":"Hi4d: 4d instance segmentation of close human interaction","cited_arxiv_id":null,"evidence_quote":"Hi4D, a close-contact interaction dataset that HHI-Assist contrasts with as the first marker-based motion-capture dataset for physical assistance."}],"review_version":1}