{"id":"d027afdc-2f47-4ace-8797-c2cc64eb4e3e","arxiv_id":"2505.14866","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UPTor couples 3D pose dynamics and trajectory prediction into one non-autoregressive transformer using a translation and rotation normalization, and adds the DARKO navigation dataset.","lead":"A robotics paper introduces UPTor, a compact transformer that predicts a person's full 3D body poses and walking path together in a global frame from a short observed clip. It also contributes DARKO, a new navigation-focused motion dataset recorded from a mobile robot's viewpoint.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DARKO and CMU-Mocap test splits share subjects with the training set, so Table II does not yet support the claimed navigation accuracy across datasets.","rationale":"I agree with the reader's overall CONDITIONAL assessment and with the secondary concern about the unstated delta in Eq. (2): no value or ablation is provided, and a noisy heading estimate for near-stationary or turning sequences could rotate every predicted global trajectory. However, I do not think that is the single most load-bearing threat to the central claim. That threat is the train/test subject overlap in the two navigation-specific datasets. The method's internal logic--translation to the root, z-axis rotation, GAT spatial embedding, non-autoregressive transformer--is coherent, and the H3.6M numbers provide some held-out-subject evidence, but those numbers are for largely static activities. Since the paper markets both a new navigation dataset and navigation-specific performance, Table II is the key evidence for the headline claim, and that table is produced under splits that allow the model to recognize test subjects from training. A subject-disjoint re-split would determine whether the reported margins are real. Until that is done, the paper should remain CONDITIONAL; no evidence here moves the verdict in either direction.","tokens_in":11412,"tokens_out":7938,"duration_ms":72248,"concrete_test":"Re-run the DARKO and CMU-Mocap experiments with strictly subject-disjoint splits. For DARKO, hold out two complete actors for testing and train on the remaining 15, or use leave-one-subject-out cross-validation over all 17 actors; for CMU-Mocap, partition by subject ID rather than by trial so no subject appears in both training and test. Recompute Table II ADE/FDE for UPTor and STPOTR under these splits. If UPTor's margins persist, the leakage concern is resolved; if margins shrink or reverse, the claim of accurate navigation prediction across DARKO and CMU-Mocap must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the evaluation protocol for the two navigation-focused datasets, not the motion transformation itself. In Sec. IV-A, the DARKO split tests on 'all trials from one actor and at least one trail from the other 16 actors' while training on the remaining trials of those same 16 actors; for CMU-Mocap the training set is drawn 'from all sampled subjects' and the test set contains 'unseen trials of activities from these subjects.' Thus 16 of 17 DARKO subjects and essentially all CMU subjects appear in both training and test partitions. A transformer can exploit subject-specific gait, body proportions, and habitual motion style, so the ADE/FDE advantages in Table II may be inflated by identity leakage rather than caused by the unified pose-trajectory formulation. The H3.6M experiment does use held-out subjects S9/S11, but the paper itself notes that the H3.6M evaluation set is dominated by static activities such as sitting, smoking, and eating, so it does not directly test the navigation claim. The abstract's promise of accuracy 'in predicting human navigation motion across all datasets' therefore rests on exactly the two splits with subject overlap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes UPTor, a unified architecture for 3D full-body pose dynamics and root-joint trajectory prediction in a global coordinate frame. The method first normalizes each motion sequence by translating the last observed root position to the origin and rotating the motion direction to the positive x-axis (the 'motion transformation'), then uses a graph attention network for spatial joint embedding, sinusoidal spatial-temporal positional encodings, and a compact non-autoregressive transformer with shared attention to decode the entire future pose-and-trajectory sequence in one pass. The output is transformed back to global coordinates using the same translation vector and rotation angle. The authors evaluate on Human3.6M, a locomotion-filtered subset of CMU-Mocap, and a newly collected egocentric navigation dataset DARKO, and report ADE/FDE for pose joints and root trajectory, runtime, and ablations. The main claims are that the unified formulation avoids the decoupling of root and pose joints, and that the model is compact, real-time, and accurate for navigational human motion.","tokens_in":11652,"tokens_out":8156,"duration_ms":69732,"significance":"If the results hold, UPTor is a useful step for human-aware robot navigation: it offers a single non-autoregressive model for joint pose and trajectory forecasting, a transformation that is straightforward and applicable to other local pose predictors, and a new egocentric navigation dataset. The derivation is not circular: the translation vector v and rotation angle theta in Eqs. (1)-(2) are computed from observed input poses, not from ground-truth future poses, and I found no step in which the predicted output is reduced to a fitted constant. The main weaknesses are in the evaluation protocol rather than in the method's internal consistency; the navigation claims currently rest on subject-overlapping splits and on point estimates with no uncertainty quantification. Strengths of the manuscript include the clear formulation of the transformation, the explicit architectural description, and the inclusion of an ablation study.","major_comments":[{"comment":"The navigation claims rest on evaluation splits that are not subject-disjoint. For DARKO, the test set draws all trials from one actor and at least one trial from the other 16 actors, while the training set contains the remaining trials of those same 16 actors; for CMU-Mocap, training is drawn from all sampled subjects and the test set contains unseen trials of activities from these subjects. A transformer can exploit subject-specific gait, body proportions, and habitual motion style, so the ADE/FDE advantages reported in Table II may reflect identity leakage rather than the unified formulation. The H3.6M experiment does use held-out subjects S9/S11, but as the authors themselves note in Sec. IV-B, that evaluation set is dominated by static activities. I ask for subject-exclusive splits for DARKO and CMU-Mocap, or per-subject results showing that the advantage holds for the held-out actor, in addition to a navigation-focused evaluation on held-out subjects.","section":"IV-A (Datasets) and Table II"},{"comment":"The orientation-invariance step depends on a two-point heading estimate computed from root positions at T1 and T1-delta, but delta is never specified and never ablated. For near-stationary poses, for turns occurring between T1-delta and T1, or for the wavy and deviating walks emphasized in DARKO, this estimate is noisy or ill-conditioned; since every input frame is rotated by theta and the predicted global trajectory is rotated back by the same theta, the final global output inherits the heading error. The paper should state delta, report the sensitivity of ADE/FDE to delta, and consider a more robust heading estimate such as an average over a longer observed window or a fit to the observed trajectory.","section":"III-B, Eq. (2)"},{"comment":"The quantitative evidence for 'real-time and accurate' is weaker than the prose suggests. In Table I, the DLow/DMMGAN/STPOTR baseline errors are copied from [3] rather than re-measured; the inference-time column I is taken from the STPOTR paper's hardware, and the R values for baselines are scaled rather than measured. All metrics are single point estimates with no confidence intervals or repeated-seed evaluation, and several differences are small (e.g., UPTor 0.51/0.74 versus STPOTR 0.50/0.75 for ADE/FDE_Po on H3.6M). Without uncertainty quantification or re-running the baselines under the same protocol, the comparisons in Table I do not fully support the claimed accuracy and runtime advantage over STPOTR.","section":"Table I and Sec. IV-B"}],"minor_comments":[{"comment":"The loss normalization appears incorrect: the sum runs from T1+1 to T1+T2, which contains T2 terms, but the prefactor is written as 1/(T2-T1-1). This should presumably be 1/T2, or 1/(T2-1) if the first predicted frame is handled differently; please correct and verify.","section":"III-B (Loss)"},{"comment":"The text contains small typos: 'at least one trail' should be 'at least one trial,' and 'such as including slow walking' should be 'such as slow walking.' Please also define T1 and T2 once in Sec. III-A and use them consistently in the loss expression.","section":"IV-A (Datasets)"},{"comment":"Figure 2 is very dense, and the decoder query initialization (repeating the encoder output for the last input pose) is not clearly visible; please enlarge the figure, add a legend for the shaded blocks, and explicitly annotate the back-transformation branch.","section":"Fig. 2"},{"comment":"The back-transformation is described only verbally; please give the explicit inverse equations (translate by -v and rotate by +theta) so that readers can verify that the global-coordinate output is correctly recovered.","section":"III-B (Motion Transformation)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper fits the robotics venue, but I would not accept it without a subject-disjoint evaluation of the DARKO and CMU-Mocap results, since the current protocol allows identity leakage. The baseline reproducibility concern in Table I is also worth conveying to the authors. I saw no citation or novelty problem; the manuscript builds transparently on POTR and STPOTR."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the motion transformation is a genuinely neat way to keep the root joint inside the pose sequence, and DARKO is a real addition to the robotics dataset pool. But the paper's headline claim — accurate navigation prediction across all datasets — rests on two splits where subjects overlap between train and test, so I read Table II as promising but not yet convincing.\n\nWhat's new: coupling pose and trajectory by translating the whole sequence so the root at T1 is at the origin, then rotating by the heading computed over a small delta, turns the problem into one non-autoregressive transformer. That differs from STPOTR's decoupled pose/trajectory branches and DMMGAN's more expensive setup. The model is compact (23M vs 43M params) and fast (17 ms). DARKO, with 17 subjects doing navigation-related actions, recorded from a robot's egocentric view, is useful. Table III is a good check: the model is invariant to translation and rotation when the transformation is on, and clearly breaks without it. Table IV's component ablation is standard but fine.\n\nWhere I'd push back: the DARKO split selects all trials of one actor and at least one trial from the other 16 for testing, with the remaining trials of those same 16 in training; CMU-Mocap is similar, unseen trials of activities from training subjects. That is identity leakage. A transformer can exploit body proportions and habitual gait, so the ADE/FDE gains in Table II may be inflated. The H3.6M experiment uses held-out subjects but, as the paper notes, is dominated by static activities, so it doesn't test the navigation claim either. Table I's baseline numbers are copied from STPOTR; only UPTor and STPOTR were run on the authors' hardware, so DLow/DMMGAN comparisons are not apples-to-apples. There are no confidence intervals or multiple seeds. And the delta in Eq. (2) is never specified or ablated; for near-stationary or turning motions, a two-point heading is unstable.\n\nNone of this is fatal. The architecture is coherent and the transformation is a substantively useful idea. What's missing is evaluation hygiene. I'd send it to review, but the revision needs subject-disjoint splits for DARKO and CMU, re-run or clearly caveated baselines, and some measure of variance. If code and data ship, that would change how much weight I put on the numbers.","headline":"Useful coupling of pose and trajectory via a motion transformation and a new navigation dataset, but the evaluation leaks subjects in both navigation-focused splits.","tokens_in":12187,"tokens_out":2872,"would_cite":true,"duration_ms":25643,"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":"UPTor claims a single non-autoregressive transformer can predict full-body pose and global trajectory together in real time using a motion-transformation canonical frame.","keywords":["human motion prediction","trajectory prediction","full-body pose forecasting","non-autoregressive transformer","graph attention network","human-robot interaction","mobile robot navigation","egocentric navigation dataset"],"falsifier":"Take a held-out set of sequences in which the person stands still for the full observation window and only starts moving after prediction time; if the two-point heading from Eq. (2) is used, the predicted global trajectory should show a large rotation error caused by the unstable angle, and this error should shrink when the heading is replaced by a longer-window or future-consistent estimate.","tokens_in":11218,"feed_emoji":"🤖","tokens_out":5020,"duration_ms":40686,"temperature":0.7,"pith_summary":"UPTor claims that full-body pose dynamics and the global trajectory of a walking person can be predicted by a single non-autoregressive transformer, provided the whole motion sequence is first normalized into a canonical coordinate frame. The normalization, called motion transformation, translates the last observed root position to the origin and rotates the observed motion direction onto the positive x-axis, so the model learns pose and root movement as one coupled prediction rather than two separate tasks. The paper reports that this compact model matches or improves on the decoupled STPOTR baseline on Human3.6M while using roughly half the transformer parameters, runs in 17 ms, and gives lower pose and trajectory errors on locomotion-heavy CMU-Mocap data and on the new DARKO robot-egocentric navigation dataset. A reader should care because real-time full-body forecasting in global coordinates is what a mobile robot needs to plan safe, socially aware navigation around people.","feed_headline":"UPTor predicts full-body pose and walking path in one 17 ms pass","feed_subtitle":"Half the parameters of the decoupled baseline, better accuracy on navigation data, built for robot planning.","key_machinery":"The load-bearing object is the motion transformation: a per-sequence normalization that subtracts the root joint position at the last observed frame from every frame and then rotates all frames by the angle between the root's recent heading and the positive x-axis. This places the observed motion and the target future in one canonical coordinate frame, so the same transformer tokens can represent pose dynamics and global root movement together, and the global prediction is recovered by applying the inverse rotation and translation. Around this sit the GAT spatial embedding, which encodes skeleton adjacency, and the non-autoregressive transformer with causal masked temporal self-attention, relative position representations, and a shared-attention decoder, which together decode the whole output horizon in a single step.","core_discovery":"The central claim is that separating the root joint from the remaining joints, as prior decoupled pose-and-trajectory methods do, is unnecessary: a transformer can output every joint including the root in global coordinates if the training sequences are first brought into a shared canonical frame. The motion transformation does this by translating each sequence so the root at the last observed frame sits at the origin and rotating the whole sequence so the root's recent motion direction points along the positive x-axis; prediction happens in that frame and the result is rotated and translated back. With a graph attention network supplying spatial skeleton structure and a non-autoregressive decoder predicting the full horizon in one pass, the resulting UPTor model achieves 0.51/0.74 m pose ADE/FDE and 0.12/0.25 m trajectory ADE/FDE on Human3.6M across all 15 actions, and 0.39/0.55 m pose and 0.13/0.26 m trajectory errors on DARKO, where it beats STPOTR. It does this with 23.2 million transformer parameters versus STPOTR's 43.3 million and a 17 ms forward pass, and the ablation shows the transformation makes trajectory predictions invariant to random global translations and rotations.","pith_inferences":["The motion transformation is stated to be applicable to any local pose prediction method, so the same canonical-frame trick could be grafted onto root-relative pose forecasters to give them global trajectory output without redesigning their decoders.","The heading used for orientation normalization is a two-point estimate over an unspecified interval delta; if that interval is not ablated, the method's claimed rotation invariance is only as strong as this estimate for near-stationary or turning motion.","The egocentric DARKO dataset could support richer contextual prediction, such as conditioning on the robot's own motion or static scene layout, since the robot's onboard sensor provides exactly the perspective a navigating robot needs.","A natural stress test is multi-person prediction, since the paper's single-person formulation would need to resolve which person's root anchors the canonical frame when several people are in view."],"forward_implications":["A single forward pass of about 17 ms yields both a full-body pose sequence and a global trajectory, making the model usable as a real-time perception module on a mobile robot.","Because pose and trajectory are learned jointly rather than decoupled, trajectory predictions can draw on pose dynamics as context, which the paper credits for its trajectory accuracy.","The motion transformation makes predictions invariant to where and in which direction a person is moving relative to the robot, as shown by the translation and rotation ablation on DARKO.","On navigation-heavy data (DARKO, locomotion subset of CMU-Mocap), the unified model reports lower pose and trajectory errors than the decoupled STPOTR baseline.","The DARKO dataset, captured from a robot's egocentric view with 17 subjects and 508 navigation trials, provides a new benchmark for human-aware navigation prediction."],"supporting_citations":[{"why":"The decoupled pose-and-trajectory baseline whose setup and metrics the paper follows and which it compares against on all three datasets.","marker":"[3]"},{"why":"Source of the non-autoregressive transformer decoder design that predicts the whole horizon in a single step.","marker":"[16]"},{"why":"Generative pose-prediction baseline used in the Human3.6M comparison.","marker":"[23]"},{"why":"Generative multi-motion baseline representing the decoupled alternative with separate trajectory and pose branches.","marker":"[20]"},{"why":"Foundation architecture for the encoder-decoder and positional encoding.","marker":"[25]"},{"why":"Graph attention network that encodes skeleton structure as spatial embeddings.","marker":"[39]"},{"why":"Relative position representations used for temporal self-attention weighting that emphasizes neighboring frames.","marker":"[40]"},{"why":"Absolute 3D pose estimator used to collect the DARKO dataset from the robot's perception stack.","marker":"[35]"},{"why":"Main public benchmark dataset used for training and evaluation.","marker":"[26]"},{"why":"Second public dataset, filtered to locomotion activities for evaluation.","marker":"[27]"}],"fun_headline_variants":["UPTor unifies pose and trajectory prediction in one pass","17 ms to predict both pose and walking path","UPTor: half the parameters, better navigation prediction","Unified motion transformer for real-time robot navigation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the heading measured between the last two input frames is a reliable reference direction for the whole prediction, which can fail when the person is nearly standing still, turning, or moving off the ground plane.","fun_headline_variants_meta":{"raw":{"variants":["UPTor unifies pose and trajectory prediction in one pass","17 ms to predict both pose and walking path","UPTor: half the parameters, better navigation prediction","Unified motion transformer for real-time robot navigation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2410,"prompt_tokens":1011,"completion_tokens":1399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1335}},"tokens_in":627,"tokens_out":1399,"duration_ms":11184,"temperature":1.0,"reasoning_tokens":1335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:27:56.399080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of sequences in which the person stands still for the full observation window and only starts moving after prediction time; if the two-point heading from Eq. (2) is used, the predicted global trajectory should show a large rotation error caused by the unstable angle, and this error should shrink when the heading is replaced by a longer-window or future-consistent estimate.","supporting_citations":[{"cited_title":"STPOTR: Simultaneous Human Trajectory and Pose Prediction Using a Non-Autoregressive Transformer for Robot Follow-Ahead","cited_arxiv_id":null,"evidence_quote":"The decoupled pose-and-trajectory baseline whose setup and metrics the paper follows and which it compares against on all three datasets."},{"cited_title":"Pose transformers (POTR): Human motion prediction with non- autoregressive transformers","cited_arxiv_id":null,"evidence_quote":"Source of the non-autoregressive transformer decoder design that predicts the whole horizon in a single step."},{"cited_title":"Dlow: Diversifying latent flows for diverse human motion prediction","cited_arxiv_id":null,"evidence_quote":"Generative pose-prediction baseline used in the Human3.6M comparison."},{"cited_title":"DMMGAN: Diverse Multi Motion Prediction of 3D Human Joints using Attention-Based Gener- ative Adversarial Network","cited_arxiv_id":null,"evidence_quote":"Generative multi-motion baseline representing the decoupled alternative with separate trajectory and pose branches."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Foundation architecture for the encoder-decoder and positional encoding."},{"cited_title":"Graph Attention Networks","cited_arxiv_id":null,"evidence_quote":"Graph attention network that encodes skeleton structure as spatial embeddings."},{"cited_title":"Self-Attention with Rela- tive Position Representations","cited_arxiv_id":null,"evidence_quote":"Relative position representations used for temporal self-attention weighting that emphasizes neighboring frames."},{"cited_title":"MeTRAbs: metric-scale truncation-robust heatmaps for absolute 3d human pose estimation","cited_arxiv_id":null,"evidence_quote":"Absolute 3D pose estimator used to collect the DARKO dataset from the robot's perception stack."},{"cited_title":"Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments","cited_arxiv_id":null,"evidence_quote":"Main public benchmark dataset used for training and evaluation."},{"cited_title":"Accessed: 01-09-","cited_arxiv_id":null,"evidence_quote":"Second public dataset, filtered to locomotion activities for evaluation."}],"review_version":1}