{"id":"a7514b7e-8719-4e96-86dc-d89669071b04","arxiv_id":"2507.14694","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ProbHMI forecasts 3D human motion by predicting Gaussian distributions over latent pose codes, which enables uncertainty quantification and diverse sampling with a compact GRU-based model.","lead":"To work safely around people, robots need to anticipate human movements. This paper introduces ProbHMI, which uses invertible networks to turn poses into a simple latent space and then forecasts a probability distribution of future motions, giving explicit uncertainty estimates. On two standard benchmarks it reports strong accuracy with a far smaller model than most baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's uncertainty claim depends on a 'quantile' that is never defined for a multivariate sequence, and the §V-D calibration check validates against a mirrored empirical proxy rather than observed coverage; this leaves the central calibration claim unverified.","rationale":"The strongest part of the paper is the invertible construction: the GCN coupling layer (Eq. 4) is genuinely invertible, and the idea of forecasting a parametric distribution in a transformed space is coherent. The accuracy results are promising and the parameter count is attractive. My concern is not about the architecture but about what the uncertainty numbers mean. The loss (Eq. 5) is a diagonal-Gaussian NLL in the latent space, so the predicted distribution is a high-dimensional product of Gaussians. To speak of a 'quantile' of such a distribution one must specify whether it is a marginal quantile, a Mahalanobis or chi-square quantile, or a joint CDF. The paper uses one scalar q for all frames and dimensions and never gives this definition. The same issue transfers to the sequence-level sampling schedule described in Section IV-E2. The calibration experiment then compares these ambiguous 'predicted quantiles' to an empirical quantile obtained by mirroring a distance-sorted set of similar-past futures; that proxy is not the true conditional distribution, and the mirroring step forces symmetry. Because the paper's headline claim is calibrated uncertainty, this is where the argument is least secure. The reader's verdict already conditions acceptance on better calibration evidence; my check would make that condition concrete. Therefore I do not move the verdict.","tokens_in":11461,"tokens_out":10135,"duration_ms":133934,"concrete_test":"Rerun the §V-D evaluation without the mirrored 'similar past motions' proxy and with an explicit quantile definition. For each test sequence, compute the standardized latent residual (z_true − μ̂)/σ̂ for every future frame and latent dimension under the predicted diagonal Gaussian, pool the residuals, and compare their empirical CDF to N(0,1); then measure empirical coverage of the predicted q-level latent hyperrectangles for q ∈ {0.25, 0.40, 0.45, 0.50} (and ideally 0.05 and 0.95). If coverage deviates from q by more than about 5 points, or if the PIT histogram is non-uniform, the calibration claim is unsupported. Also report the actual joint CDF of the generated 'Pq' trajectories as a check that the labels denote true quantiles.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ProbHMI's central claim of 'validated uncertainty calibration' rests on two unsecured steps. First, Section IV-E defines frame-level uncertainty as 'the quantile associated with the latent code' and then applies the same quantile to all frames and latent dimensions when sampling, but a scalar quantile is not well-defined for a multivariate diagonal Gaussian: the point whose coordinates are each the q-th marginal quantile has joint CDF q^d (or central-region mass (1-2q)^d), not q. The paper gives no formula for a joint or sequence-level quantile, so the labels P50, P45, P40, and P25 in Figures 4 and 5 do not correspond to a stated probability. Second, the calibration evaluation in Section V-D constructs an empirical distribution from test sequences with similar past motions, sorts the subsequent motions by distance, and 'mirror[s] the empirical distribution to match the form of the predicted distribution.' That symmetrization removes the very skew the Gaussian model cannot represent and aligns the proxy with the model's median, so low ADE and FDE in Figure 5 is not evidence that predicted quantiles match true quantiles. The model could be substantially miscalibrated and still pass this test. Since the headline contribution over prior work is explicit, calibrated uncertainty, this gap is load-bearing; the accuracy numbers alone do not support it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ProbHMI, a two-module framework for 3D human motion forecasting. A part-aware invertible network (PTM) maps poses to a disentangled latent space, and a single GRU (PFM) predicts per-frame factorized Gaussian parameters in that space. The paper claims state-of-the-art accuracy on Human3.6M and HumanEva-I for both deterministic and diverse prediction, validated uncertainty calibration via an empirical quantile evaluation, and efficient sampling with few samples. The central methodological novelty is the explicit probabilistic formulation enabled by invertible transformations.","tokens_in":11778,"tokens_out":5083,"duration_ms":61291,"significance":"If the uncertainty-calibration claim were fully supported, this would be a valuable contribution: ProbHMI uses only 0.36M parameters, achieves competitive ADE and the best FID on Human3.6M, and provides a concrete mechanism for extracting confidence values from a generative motion model. The part-aware coupling layer and the explicit latent-space Gaussian forecasting are sensible design choices, and the ablations support the benefit of the part-aware modules. However, the headline contribution over prior work is 'validated uncertainty calibration,' and the current evaluation does not establish it: the quantile used for sampling is not defined for a multivariate Gaussian, and the empirical calibration protocol in Section V-D is not a coverage test. The accuracy and efficiency results are also weakened by an uncontrolled sampling comparison in Section V-E. These issues are fixable but require additional experiments.","major_comments":[{"comment":"In Section IV-E and Equation (3), the uncertainty for a sequence is defined by applying the 'same quantile to all frames' during sampling, and Figure 5 labels results as P50, P45, P40, P25. However, for the factorized diagonal Gaussian over the latent code (Eq. 5), a scalar quantile is not well-defined: if each coordinate is set to its q-th marginal quantile, the joint probability mass is q^d (or (1-2q)^d for a central region), not q. No definition of a joint or sequence-level quantile is provided, so the reported percentile labels do not correspond to a stated probability. This is load-bearing because the central claim is calibrated uncertainty.","section":"Section IV-E, Eq. (3), Eq. (5), Figure 5"},{"comment":"Section V-D's calibration evaluation does not measure calibration. The authors build an empirical distribution from test sequences with sufficiently similar past motion, sort subsequent motions by distance, and then 'mirror the empirical distribution to match the form of the predicted distribution.' This mirroring removes skew in the empirical proxy and aligns the proxy's median with the model's symmetric median, so the low ADE/FDE values in Figure 5 are expected even for a miscalibrated model. The evaluation should instead report coverage rates or probability integral transform values for the predicted latent/pose distributions, and should not symmetrize the target distribution. The paper's own caveat about 'approximations in the empirical ground truth' does not fix this issue.","section":"Section V-D, Figure 5"},{"comment":"The sampling-efficiency comparison in Section V-E and Table V is not controlled: ProbHMI uses Poisson-Disk Sampling while DLow, GSPS, and HumanMAC use a vanilla sampling schedule. The reported advantage therefore conflates the sampling algorithm with the model. To support the claim that ProbHMI estimates the future distribution with few samples, all methods should use the same sample-selection strategy, or the paper should ablate the sampler within ProbHMI.","section":"Section V-E, Table V"},{"comment":"Equation (5) writes the negative log-likelihood as (1/K) * sum [log(Sigma_i) + (z_i - z_hat_i)^2 / (2 (Sigma_i)^2)], and the text states that Sigma denotes variance. If Sigma is the variance, the quadratic term should be divided by 2*Sigma and the log term should be log(Sigma) without the additional square; if Sigma is the standard deviation, the notation should say so explicitly. As written, the optimized objective is not the Gaussian negative log-likelihood stated, which matters for a paper whose contribution is probabilistic modeling.","section":"Section IV-D, Eq. (5)"}],"minor_comments":[{"comment":"The loss LN is described as minimizing a KL divergence but the expression is the negative log-likelihood of the latent code under the standard Gaussian plus a Jacobian term; please clarify the relationship between this objective and the stated KL goal.","section":"Section IV-D, Eq. (7)"},{"comment":"'auto-aggressive' should be 'autoregressive.'","section":"Section IV-E"},{"comment":"The visualization captions use opacity/quantile mappings without defining how opacity maps to the quantile levels; please add a legend or explicit mapping.","section":"Figures 3 and 4"},{"comment":"The hyperparameters for the empirical quantile evaluation (distance threshold 0.5 in Figure 4, the minimum of 50 pseudo-futures, and the mirroring procedure) are described only in prose; a precise algorithmic description would improve reproducibility.","section":"Section V-D"},{"comment":"The conclusion contains a grammatical error: 'incorporating stronger motion priors ... may holds promise' should be 'may hold promise.'","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper's advertised contribution is calibrated uncertainty, but the current quantile definition and calibration protocol are not sound. I recommend requiring the authors to (1) define a proper multivariate or sequence-level quantile, (2) evaluate calibration with coverage or PIT-based metrics without symmetrizing the target distribution, and (3) control the sampling algorithm in the efficiency comparison. If these are addressed, the paper could be a solid contribution; as it stands, the central claim is unverified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead the ProbHMI paper. The short version: the accuracy story is legit and the invertible-network idea is worth taking seriously; the uncertainty calibration story does not survive contact with the paper's own equations. If you cite it, cite it for the architecture and the deterministic/diverse numbers, not for calibrated quantiles.\n\nWhat's new: mapping poses through a part-aware invertible network (GCN-based additive coupling) and forecasting a diagonal Gaussian in that latent space with a single GRU. That's a clean, compact formulation, and the 0.36M-parameter model beating much larger baselines on ADE and FID on Human3.6M is a real result. The ablation against standard NICE is the right control and it shows the part-aware design is doing work. The deterministic MAE comparison against ResGRU is also persuasive: same GRU minus the PTM, so the latent-space forecasting is what improves accuracy.\n\nSoft spots, in order of size. (1) The quantile is never defined. Section IV-E says frame-level uncertainty is 'the quantile associated with the latent code' and then says they apply the same quantile to all frames, but there is no formula mapping a requested probability to a covariance scale or a sample. For a multivariate diagonal Gaussian, coordinate-wise quantiles are not a joint quantile; P50/P45/P40/P25 in Figures 4 and 5 are labels without a defined probability. (2) The calibration check in Section V-D constructs the 'true' distribution from similar test sequences, then mirrors it 'to match the form of the predicted distribution.' That symmetrization removes exactly the skew a symmetric Gaussian cannot capture, so low ADE and FDE there is not evidence of calibrated uncertainty. The authors do acknowledge the proxy is approximate, but the conclusion 'predictions still reflect movement trends' is weaker than the abstract's claim of validated calibration. (3) The sampling-efficiency table gives ProbHMI Poisson-Disk sampling while baselines get vanilla sampling; that is not an apples-to-apples comparison. (4) Minor: no code, no error bars, and Sigma is used for both variance and std in places.\n\nThis is a paper with a good core idea and strong benchmark numbers, but the central uncertainty claim is not yet evidenced. It deserves a serious referee, not a desk reject — the architecture and accuracy results are enough to warrant a careful revision, and the uncertainty evaluation needs rethinking, e.g., proper calibration plots or conformal coverage. I'd send it to review but tell the authors to fix the quantile definition and the calibration protocol.","headline":"Solid accuracy results and a clean invertible-latent formulation, but the headline uncertainty calibration claim rests on an undefined multivariate quantile and a mirrored proxy that cannot validate it.","tokens_in":12251,"tokens_out":1963,"would_cite":true,"duration_ms":23975,"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":"ProbHMI shows that 3D human motion forecasting can be formulated as explicit probability distributions over future poses, using invertible networks to map poses into a continuous latent space, so that accuracy, diversity, and calibrated…","keywords":["3D human motion forecasting","uncertainty quantification","invertible networks","probabilistic forecasting","diverse motion prediction","GRU","part-aware coupling layers"],"falsifier":"A direct calibration test would settle the claim: on a held-out split, for each predicted quantile q in {0.25, 0.40, 0.45, 0.50}, compute the empirical frequency with which the true future pose falls inside the predicted q-level region; systematic mismatch, such as the claimed 25th percentile actually containing half of the true futures, would falsify the uncertainty-calibration claim.","tokens_in":11303,"feed_emoji":"🤖","tokens_out":10361,"duration_ms":110762,"temperature":0.7,"pith_summary":"The paper sets out to show that 3D human motion forecasting can be made explicitly probabilistic—returning a distribution over future poses instead of a bag of samples—without sacrificing predictive accuracy. Its proposed model, ProbHMI, maps each observed pose through an invertible network into a continuous latent space, then uses a single GRU to forecast the mean and variance of the next latent code; sampling from that distribution produces diverse futures, and the density or quantile of the sample provides an uncertainty estimate. This matters for human-robot collaboration, where a robot needs not just a most-likely motion but a confidence measure to plan safely. The authors report the approach outperforms prior diverse-forecasting methods on average displacement error and Fréchet inception distance on Human3.6M while using 0.36M parameters, and that its quantile-based predictions remain close to the ground truth at every tested percentile.","feed_headline":"Invertible latent maps give 3D motion forecasts calibrated uncertainty","feed_subtitle":"ProbHMI turns pose forecasting into explicit probability distributions that a robot can query for risk-aware decisions.","key_machinery":"The load-bearing object is the Pose Transformation Module (PTM), an invertible network built from graph-convolution-based additive coupling layers, a topology-preserving variant of the NICE architecture. Each coupling layer splits the skeleton graph into two parts, for example upper and lower body, passes one part through unchanged, and adds to the other a graph-convolutional function of the first; because the map is bijective, the exact log-likelihood of a pose is available through the change-of-variables formula. This lets the Pose Forecasting Module, implemented as a single GRU with part-aware heads, treat the next latent code as a draw from a factorized Gaussian with predicted mean $\\hat{z}$ and diagonal covariance $\\Sigma$, rather than as an implicit sample. Invertibility is what converts an intractable pose distribution into a parametric latent distribution while still guaranteeing that sampled latent codes decode into plausible poses.","core_discovery":"The paper's central claim is that explicit probability distributions over future 3D poses can be learned without giving up accuracy or diversity: map each pose through a topology-preserving invertible network into a continuous latent space, forecast the mean and diagonal covariance of the next latent code with a single GRU, and decode samples back into poses. Reported results on Human3.6M and HumanEva-I indicate that this one design, with 0.36M parameters, achieves the lowest ADE (0.364) and FID (0.646) among compared diverse-forecasting methods on Human3.6M, improves deterministic mean angle error across all horizons, and degrades only mildly when the number of samples is cut from 50 to 5 (ADE increases 6.04% versus 50.82% for DLow). The same explicit distribution supplies frame- and sequence-level quantiles, which the authors validate against an empirical quantile constructed from test motions with similar histories.","pith_inferences":["Because the interface is a full predictive distribution, the framework could plug directly into conformal prediction or risk-sensitive control loops that consume quantiles; the paper does not build that layer.","The factorized-Gaussian assumption in latent space could be tested action-by-action: motions with genuinely multi-modal futures, such as walking versus turning, may need a mixture or flow-based posterior, a natural extension.","Calibration could be checked more directly on data with repeated takes of the same scripted motion, where the true future distribution is known, instead of the similar-past-motion proxy used in the paper's empirical quantile evaluation.","If invertibility is the source of the gain, replacing the single GRU with a more expressive dynamics model while keeping the invertible pose map should improve accuracy further without changing the uncertainty interface."],"forward_implications":["A robot or autonomous system can query the forecast distribution directly, reading off a probability density or quantile per frame and per sequence, instead of approximating uncertainty by sampling many trajectories and measuring spread.","Because the predicted distribution is parametric, few samples are needed to cover plausible futures: at five samples instead of fifty, ProbHMI's ADE rises only 6.04%, while DLow's rises 50.82%.","Forecasting in the invertible latent space is itself beneficial for deterministic accuracy: ProbHMI beats the same recurrent architecture operating directly on poses, with 0.86 versus 1.18 mean angle error at 400 ms.","The part-aware, topology-preserving design contributes to the gain: removing part-aware prediction or substituting the standard NICE invertible network increases ADE and FID.","Explicit likelihood also enables alternative sampling schedules, such as the Poisson-disk sampling used in the five-sample evaluation, a tool not available to implicit generative models."],"supporting_citations":[{"why":"Supplies the DLow diverse-forecasting baseline that ProbHMI compares against on Human3.6M and HumanEva-I, including the five-sample efficiency comparison.","marker":"[7]"},{"why":"Baseline GSPS for smooth diverse motion prediction, used in the main diverse tables, qualitative comparison, and five-sample evaluation.","marker":"[8]"},{"why":"Baseline HumanMAC compared in the diverse, qualitative, and five-sample experiments.","marker":"[14]"},{"why":"Deterministic baseline ResGRU, whose recurrent architecture matches ProbHMI minus the invertible module, isolating the gain from latent-space forecasting.","marker":"[18]"},{"why":"Baseline Motron used in both diverse and deterministic evaluation and in qualitative comparisons.","marker":"[31]"},{"why":"NICE, the invertible-network architecture whose additive coupling layers ProbHMI adapts into graph-convolution-based part-aware layers.","marker":"[43]"},{"why":"Human3.6M dataset used for the diverse and deterministic evaluations.","marker":"[48]"},{"why":"HumanEva-I dataset used for the diverse evaluation.","marker":"[49]"}],"fun_headline_variants":["Invertible networks give 3D motion forecasts explicit uncertainty","ProbHMI: motion forecasts with calibrated confidence for robots","Forecast human motion as distributions, not point estimates","Invertible latent dynamics yield calibrated 3D pose predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true distribution of future motions is well captured by a factorized Gaussian with diagonal covariance in the invertibly mapped latent space, so that predicted variances and quantiles carry the same meaning in pose space.","fun_headline_variants_meta":{"raw":{"variants":["Invertible networks give 3D motion forecasts explicit uncertainty","ProbHMI: motion forecasts with calibrated confidence for robots","Forecast human motion as distributions, not point estimates","Invertible latent dynamics yield calibrated 3D pose predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000345,"raw_usage":{"total_tokens":1847,"prompt_tokens":852,"completion_tokens":995,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":927}},"tokens_in":468,"tokens_out":995,"duration_ms":8918,"temperature":1.0,"reasoning_tokens":927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:49:24.336216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct calibration test would settle the claim: on a held-out split, for each predicted quantile q in {0.25, 0.40, 0.45, 0.50}, compute the empirical frequency with which the true future pose falls inside the predicted q-level region; systematic mismatch, such as the claimed 25th percentile actually containing half of the true futures, would falsify the uncertainty-calibration claim.","supporting_citations":[{"cited_title":"Dlow: Diversifying latent flows for diverse human motion prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the DLow diverse-forecasting baseline that ProbHMI compares against on Human3.6M and HumanEva-I, including the five-sample efficiency comparison."},{"cited_title":"Generating smooth pose se- quences for diverse human motion prediction,","cited_arxiv_id":null,"evidence_quote":"Baseline GSPS for smooth diverse motion prediction, used in the main diverse tables, qualitative comparison, and five-sample evaluation."},{"cited_title":"HumanMAC: Masked Motion Completion for Human Motion Prediction","cited_arxiv_id":"2302.03665","evidence_quote":"Baseline HumanMAC compared in the diverse, qualitative, and five-sample experiments."},{"cited_title":"On human motion prediction using recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Deterministic baseline ResGRU, whose recurrent architecture matches ProbHMI minus the invertible module, isolating the gain from latent-space forecasting."},{"cited_title":"Motron: Multimodal proba- bilistic human motion forecasting,","cited_arxiv_id":null,"evidence_quote":"Baseline Motron used in both diverse and deterministic evaluation and in qualitative comparisons."},{"cited_title":"Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments,","cited_arxiv_id":null,"evidence_quote":"Human3.6M dataset used for the diverse and deterministic evaluations."},{"cited_title":"Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion,","cited_arxiv_id":null,"evidence_quote":"HumanEva-I dataset used for the diverse evaluation."}],"review_version":1}