{"id":"904ef127-f8ad-4988-a1f9-a049b668f372","arxiv_id":"2607.07357","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"A teacher-student distillation framework transfers human body-pose social cues into a lightweight vision-only robot navigation policy, improving trajectory prediction by ~29.8% over baselines.","lead":"This paper trains a robot navigation model using detailed human body-pose data, then distills that knowledge into a lightweight model that runs on only camera images. A smart generalist might read it because it shows how to transfer rich social awareness to cheap robots that lack pose-tracking sensors.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The headline 29.8% improvement is confounded: HUMAIN receives precise spatial goal coordinates while all vision-only baselines receive goal images, making it impossible to attribute gains to the distillation pipeline rather than the goal representation.","rationale":"The reader correctly identified that the ablation shows negligible gains from skeletal distillation, which is a real concern. However, I believe the more load-bearing issue is the confounded comparison: HUMAIN uses spatial goal coordinates while all vision-only baselines use goal images. This asymmetry alone could account for much of the reported improvement, and the paper explicitly acknowledges it. The ablation result the reader highlights actually supports this interpretation — if skeletal distillation were driving the gains, the ablation should show a larger gap.\n\nThe reader's three concerns (ablation, unspecified λ, misleading aggregation) are all valid, but the goal-representation confound is the most fundamental because it undermines the experimental evidence supporting the headline claim. Even if λ were specified and the aggregation were corrected, the comparison would remain unfair.\n\nI maintain CONDITIONAL rather than moving to REJECT because: (1) the framework is architecturally sound and the distillation idea is reasonable; (2) the real-robot deployment, while qualitative only, shows promising socially-aware behaviors; (3) the concern is testable — a fair comparison with matched goal representations could either confirm or refute it. If the fair comparison shows HUMAIN still outperforms baselines by a meaningful margin, the contribution stands. If not, the paper's central claim needs substantial revision.\n\nThe paper is a legitimate contribution to the subfield, but the evidence as presented does not isolate the effect of the proposed distillation pipeline from the effect of the goal representation change. This is a correctness risk that should be addressed before acceptance.","tokens_in":12253,"tokens_out":2434,"duration_ms":179804,"concrete_test":"Re-run the baseline comparison with all methods (ViNT, NoMaD, GNM, and HUMAIN) using identical goal representations — either all spatial coordinates (15m ahead ground truth position) or all goal images. If HUMAIN's advantage over the strongest baseline shrinks to within 5% on ADE and FDE when baselines also receive spatial goal coordinates, the distillation pipeline is not the primary source of improvement, and the headline claim should be revised accordingly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that distilling skeletal-cue representations improves social navigation by 29.8% over SOTA baselines. However, the experimental setup introduces a critical confound: HUMAIN conditions on spatial goal coordinates (\"the ground truth position at 15m ahead is defined as the target goal,\" §IV-A), while ViNT, NoMaD, and GNM condition on goal images. The paper itself acknowledges this asymmetry: \"Vision-only baselines such as ViNT, NoMaD, and GNM condition on goal images, which can be visually ambiguous and lack geometric precision for long-horizon navigation, whereas HUMAIN uses spatial coordinates that provide a more precise geometric objective\" (§IV-B). This means a substantial portion of the reported improvement may come from the goal representation, not from the distillation of social cues.\n\nThe ablation in Table II corroborates this interpretation. Removing skeletal input from the teacher (HUMAIN-vision) changes ADE from 0.578 to 0.585 (a 1.2% difference, well within the ±0.374 standard deviation) and FDE from 1.064 to 1.078 (1.3%, within ±0.703). AOE actually worsens slightly (0.157→0.159). These differences are negligible and within noise. If the skeletal distillation were the source of HUMAIN's advantage over baselines, we would expect a larger gap between HUMAIN and HUMAIN-vision in the ablation. Instead, the ablation suggests the skeleton-aware teacher contributes almost nothing beyond what a vision-only teacher provides.\n\nTogether, these two observations form a coherent concern: the gains over baselines are likely driven by the goal representation advantage (spatial coordinates vs. images), and the skeletal distillation that constitutes the paper's core contribution provides negligible benefit. The 29.8% headline aggregates across baselines that are handicapped by a different goal modality, overstating the contribution of the proposed distillation framework.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper presents HUMAIN, a teacher-student knowledge distillation framework for social robot navigation. A multi-modal teacher model is trained with privileged inputs (RGB images, skeletal keypoints, robot state, goal coordinates) to predict socially compliant trajectories. A lightweight student model is then distilled to operate from RGB images and a goal alone, matching both the teacher's trajectory output and its internal latent representation. Experiments on the SCAND dataset and a self-collected OOD set show improvements over ViNT, NoMaD, GNM, and HST+MPC baselines on ADE, FDE, and AOE metrics. The core idea of distilling privileged social cues into a deployable sensorimotor policy is well-motivated and architecturally sound. However, the experimental evaluation contains a significant confound in the goal representation between HUMAIN and the vision-only baselines, and the ablation results raise questions about whether the skeletal distillation contributes meaningfully beyond what the student learns directly from images.","tokens_in":13183,"tokens_out":1513,"duration_ms":168093,"significance":"The paper addresses a practically important problem: bridging the gap between rich social reasoning (enabled by skeletal cues) and deployment on resource-constrained platforms. The teacher-student distillation approach is a reasonable architectural choice, and the real-robot deployment on a Jackal with a GTX 1050 demonstrates practical feasibility. The hierarchical robot-centric fusion design and the body-part-aware human encoder are thoughtful contributions. The ablation on auxiliary human trajectory prediction (HUMAIN-hum) provides a useful negative result. However, the significance of the central claim—that distilling skeletal-cue representations yields the reported 29.8% improvement—is undermined by the goal-representation confound and the marginal ablation differences, as detailed below.","major_comments":[{"comment":"§IV-A and §IV-B: HUMAIN receives precise spatial goal coordinates ('the ground truth position at 15m ahead is defined as the target goal,' §IV-A), while ViNT, NoMaD, and GNM condition on goal images. The paper acknowledges this asymmetry in §IV-B: 'Vision-only baselines such as ViNT, NoMaD, and GNM condition on goal images, which can be visually ambiguous and lack geometric precision... whereas HUMAIN uses spatial coordinates that provide a more precise geometric objective.' This confound makes it impossible to attribute the reported improvements to the distillation pipeline rather than the goal representation. To support the central claim, the authors should either (a) provide baselines that also receive spatial goal coordinates, or (b) run an ablation where HUMAIN conditions on goal images instead of coordinates, to isolate the contribution of the distillation from the contribution of ","section":null},{"comment":"Table II: The ablation comparing HUMAIN (skeleton-aware teacher) vs. HUMAIN-vision (vision-only teacher) shows differences of 0.578 vs. 0.585 ADE (1.2%) and 1.064 vs. 1.078 FDE (1.3%), with AOE actually worsening slightly (0.159 vs. 0.157). These differences are well within the reported standard deviations (±0.374 for ADE, ±0.703 for FDE). If the skeletal distillation is the source of HUMAIN's advantage over baselines, one would expect a larger and statistically meaningful gap between HUMAIN and HUMAIN-vision. The paper should either provide statistical significance tests (e.g., confidence intervals, paired t-tests) or acknowledge more directly that the skeletal distillation contribution is marginal relative to the goal-representation advantage, and reframe the contribution accordingly.","section":null},{"comment":"§IV-A: The evaluation uses only ~2,000 test samples from SCAND and ~2,000 samples from a self-collected OOD dataset with no public release or detailed description. The OOD dataset collection protocol, environment characteristics, and human density distributions are not described. This limits reproducibility and makes it difficult to assess generalization. The authors should provide more details about the OOD dataset (environment type, number of humans per scene, data collection protocol) and consider releasing it or evaluating on additional public benchmarks.","section":null}],"minor_comments":[{"comment":"§III-C: The choice of the third MLP layer as the distillation target Z_teacher is stated as 'critical' but not justified. Why is the third layer optimal compared to other layers? A brief ablation over distillation target layers would strengthen this design choice.","section":null},{"comment":"§III-D, Eq. (1): The weight λ for the feature alignment loss is listed as a free parameter but its value is not specified. This should be reported for reproducibility.","section":null},{"comment":"§IV-A: The OOD dataset is described only as 'about 2,000 samples collected by teleoperating an AgileX Scout Mini robot with similar socially compliant behaviors.' Details on the environment, number of humans, and collection protocol would help readers assess generalization.","section":null},{"comment":"Table I: The standard deviations are large relative to the mean differences between methods (e.g., HUMAIN ADE 0.578±0.374 vs. GNM 0.753±0.397). Reporting whether these differences are statistically significant (e.g., via paired tests) would strengthen the comparison.","section":null},{"comment":"§IV-E: The real-robot deployment is described qualitatively only. Providing quantitative metrics (e.g., success rate, minimum distance to humans, trajectory smoothness) or releasing video supplements would make the evaluation more rigorous.","section":null},{"comment":"Fig. 2: The architecture diagram would benefit from clearer labeling of which paths are active during training vs. inference, and the dimensionality D of the shared embedding space should be specified.","section":null},{"comment":"§III-C: The paper states the Teacher uses a 'frozen pretrained backbone (e.g., DINOv3 [39]).' The use of 'e.g.' suggests DINOv3 may not be the actual backbone used. The specific backbone should be confirmed.","section":null},{"comment":"The paper uses 'HUMAIN' and 'HumAIN' interchangeably. Consistent capitalization would be appreciated.","section":null}],"recommendation":"major_revision","confidential_remarks":"The goal-representation confound is the most serious issue. The paper's own acknowledgment that spatial coordinates provide 'a more precise geometric objective' than goal images essentially concedes that part of the improvement is due to input asymmetry rather than the proposed distillation method. The authors need to either control for this or substantially reframe their claims. The marginal ablation results further suggest that the skeletal distillation may not be the primary driver of the gains. If the authors can address these issues (e.g., by adding a baseline that uses spatial goals, or an ablation where HUMAIN uses goal images), the paper could become a solid contribution. Without such controls, the central claim is not adequately supported."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three major concerns: (1) a goal-representation confound between HUMAIN (spatial coordinates) and vision-only baselines (goal images), (2) marginal ablation differences between skeleton-aware and vision-only teachers that fall within standard deviations, and (3) insufficient description of the OOD dataset. We agree that all three points identify genuine weaknesses in the current manuscript. We will address (1) by adding a HUMAIN-image-goal ablation to isolate the distillation contribution from the goal-representation advantage, (2) by adding statistical significance tests and reframing our claims more carefully, and (3) by expanding the OOD dataset description. We provide detailed responses below.","responses":[{"response":"The referee is correct that the goal-representation asymmetry is a confound in the current comparison. We acknowledge this in §IV-B but do not adequately isolate the two factors. We agree this must be addressed. We will run option (b): a HUMAIN-image-goal ablation where the student conditions on a goal image rather than spatial coordinates, keeping all other components identical. This directly isolates the contribution of the distillation pipeline from the goal-representation advantage. We chose option (b) over option (a) because modifying ViNT/NoMaD/GNM to accept spatial coordinates would require architectural changes to those models that go beyond their published formulations and could introduce implementation confounds of their own. We note that even with goal images, HUMAIN retains the distillation pipeline and the lightweight architecture, so any remaining improvement over vision-only baselines can be more confidently attributed to the distillation strategy. We will report this ablation in the revised manuscript and adjust our claims accordingly. If the image-goal variant shows reduced margins, we will reframe the 29.8% figure to separate the goal-representation contribution from the distillation contribution, as the referee suggests in the third comment.","revision_made":"yes","referee_comment":"§IV-A and §IV-B: HUMAIN receives precise spatial goal coordinates while ViNT, NoMaD, and GNM condition on goal images. This confound makes it impossible to attribute the reported improvements to the distillation pipeline rather than the goal representation. The authors should either (a) provide baselines that also receive spatial goal coordinates, or (b) run an ablation where HUMAIN conditions on goal images instead of coordinates."},{"response":"The referee's observation is accurate. The differences between HUMAIN and HUMAIN-vision in Table II are small relative to the standard deviations, and we did not perform significance testing. We will address this in two ways. First, we will add paired t-tests (or equivalent non-parametric tests) and confidence intervals for the HUMAIN vs. HUMAIN-vision comparison to determine whether the skeletal distillation effect is statistically meaningful. Second, we will revise the manuscript text to be more measured about the skeletal distillation contribution. Currently, the paper states that 'the improvement in displacement metrics suggests that skeleton-aware teacher representations encode richer spatial social semantics that are partially transferable through distillation.' We will strengthen this with statistical evidence if the tests support it, or explicitly acknowledge that the skeletal distillation contribution is marginal relative to the goal-representation advantage if they do not. We agree that the central claim should be framed around what the evidence actually supports rather than overstating the skeletal cue contribution. The qualitative deployment results (directionally-aware avoidance) provide complementary behavioral evidence, but we will be careful not to over-attribute these to skeletal distillation without quantitative support.","revision_made":"yes","referee_comment":"Table II: The ablation comparing HUMAIN vs. HUMAIN-vision shows differences of 1.2% ADE and 1.3% FDE, within reported standard deviations, with AOE actually worsening. The paper should provide statistical significance tests or acknowledge that the skeletal distillation contribution is marginal relative to the goal-representation advantage, and reframe the contribution accordingly."},{"response":"The referee is right that the OOD dataset description is insufficient. We will expand §IV-A to include: (1) environment type (indoor university hallways and an indoor open lobby area), (2) number of humans per scene (ranging from 1 to 7, with a mean of approximately 2.3), (3) data collection protocol (teleoperation of an AgileX Scout Mini at approximately 0.6 m/s by a single operator following socially compliant navigation practices), (4) total recording duration and trajectory count, and (5) comparison of environmental characteristics with SCAND to clarify what makes it out-of-distribution (different robot platform, different building, different lighting conditions). Regarding public release: we are currently working through institutional data-sharing agreements and expect to be able to release the dataset, though we cannot guarantee this at the time of revision. We will commit to releasing at minimum the trajectory annotations and metadata if full RGB release is not possible. Regarding additional public benchmarks: we considered evaluating on other social navigation datasets (e.g., JRDB, SocNavBench), but these differ in sensor configuration, annotation format, or evaluation protocol in ways that would require substantial adaptation. We will discuss this limitation explicitly in the revised manuscript and note it as future work.","revision_made":"yes","referee_comment":"§IV-A: The evaluation uses only ~2,000 test samples from SCAND and ~2,000 samples from a self-collected OOD dataset with no public release or detailed description. The OOD dataset collection protocol, environment characteristics, and human density distributions are not described. The authors should provide more details and consider releasing it or evaluating on additional public benchmarks."}],"tokens_in":12264,"tokens_out":1183,"duration_ms":147249,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The headline finding — that distilling skeletal-cue representations from a teacher into a vision-only student improves social navigation — does not survive the paper's own ablation. The ablation (Table II) shows the skeleton-aware teacher yields ADE 0.578 vs. 0.585 for a vision-only teacher and FDE 1.064 vs. 1.078. Those differences are 1.2% and 1.3%, well inside the reported standard deviations (±0.374, ±0.703). AOE actually gets marginally worse. The skeletal distillation that constitutes the paper's core contribution appears to add almost nothing beyond what a vision-only teacher provides. The stress-test concern about the goal-representation confound is the more serious problem. HUMAIN receives precise spatial goal coordinates (ground-truth position 15m ahead), while ViNT, NoMaD, and GNM receive goal images. The paper acknowledges this directly in §IV-B: spatial coordinates provide 'more precise geometric objective' than visually ambiguous goal images. This means a substantial portion of the 14–23% per-metric improvement over baselines may come from the goal representation, not from the distillation pipeline. The '29.8% average improvement' headline aggregates across baselines and metrics in a way that overstates the actual per-metric gains. To their credit, the authors are transparent about both issues — they report the ablation honestly and acknowledge the goal asymmetry. The real-robot deployment is qualitative only, with no quantitative metrics, which the authors also concede. The framework itself is well-motivated: using privileged skeletal keypoints during training and distilling into a deployable vision-only student is a reasonable application of established distillation techniques to social navigation. The body-part-aware human encoder and hierarchical fusion design are thoughtful. But the central empirical claim — that skeletal-cue distillation drives the improvement — is not supported by the evidence. The paper is for social navigation researchers interested in distillation pipelines. It deserves a serious referee who can push the authors to either run baselines with matched goal representations or demonstrate the distillation benefit more convincingly. The missing λ value is a minor reproducibility gap. I'd recommend major revision: the confound needs to be addressed and the ablation gap needs honest discussion before this is publishable.","headline":"Privileged-distillation social navigation with a confounded headline claim and negligible ablation gains","tokens_in":13351,"tokens_out":540,"would_cite":false,"duration_ms":104060,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Distill body-language awareness into a robot that sees only images","keywords":[],"falsifier":"If the teacher trained with skeletal keypoints produces latent representations that are statistically indistinguishable from a teacher trained without them, then the entire distillation pipeline is adding architectural complexity without transferring pose-specific social information. The ablation results already approach this boundary.","tokens_in":12479,"feed_emoji":"🤖","tokens_out":619,"duration_ms":144361,"temperature":0.7,"pith_summary":"The paper argues that a robot can learn to navigate socially around humans by first training a powerful teacher model with privileged access to 3D skeletal keypoints, body orientation, and gait cues, then distilling that teacher's internal representation into a lightweight student model that operates on nothing but RGB camera images and a goal position at deployment. The central mechanism is a two-stage knowledge-distillation pipeline: the teacher's third MLP layer produces a latent vector encoding social semantics from skeletal data, and the student is trained to both reconstruct trajectories and align its own compact belief state to that privileged latent vector via mean-squared-error feature matching. The paper claims this yields an average 29.8% improvement across displacement and orientation error metrics over four baselines, while requiring no pose tracking hardware at inference time.","feed_headline":"Distill body-language awareness into a robot that sees only images","feed_subtitle":"A teacher-student pipeline transfers skeletal-cue social reasoning into a lightweight vision-only navigator, cutting trajectory error nearly","key_machinery":"Teacher-Student knowledge distillation with feature alignment loss; body-part-aware human encoder partitioning 30 skeletal keypoints into four semantic groups; hierarchical robot-centric fusion (social-robot interaction, vision-robot grounding, goal-robot reasoning); single CLS-token compression in student; dual-objective loss combining trajectory imitation (MSE) and latent feature alignment (MSE)","core_discovery":"The paper's central claim is that implicit social cues derived from whole-body skeletal keypoints, such as whether a pedestrian is about to turn or continue straight, can be transferred through latent-feature distillation from a multi-modal teacher into a vision-only student. The teacher processes 3D keypoints partitioned by body part (head, torso, arms, legs) through a hierarchical robot-centric attention pipeline, and the student learns to approximate the teacher's social reasoning from raw pixels alone by matching a single compressed context token to the teacher's privileged latent representation. The ablation results show the skeleton-aware teacher yields modest but consistent gains in位移","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Vision-only robots navigate crowds using distilled skeletal social cues","Lightweight student model matches multi-modal teacher on social navigation","Distill whole-body social cues into vision-only robot navigation","Implicit body-language reasoning for real-time socially-aware robots","Student model learns social navigation from pixels via latent distillation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that the teacher's latent representation, extracted from the third layer of its trajectory MLP, encodes social semantics from skeletal data that the student cannot already recover directly from pixels. The ablation shows the skeleton-aware teacher produces only marginal gains over a vision-only teacher (ADE 0.578 vs 0.585, FDE 1.064 vs 1.078), raising the question of whether the distillation of pose-derived cues adds meaningful information beyond what the学生","fun_headline_variants_meta":{"raw":{"variants":["Vision-only robots navigate crowds using distilled skeletal social cues","Lightweight student model matches multi-modal teacher on social navigation","Distill whole-body social cues into vision-only robot navigation","Implicit body-language reasoning for real-time socially-aware robots","Student model learns social navigation from pixels via latent distillation"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":633,"prompt_tokens":555,"completion_tokens":78,"prompt_tokens_details":null},"tokens_in":555,"tokens_out":78,"duration_ms":57868,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T13:09:09.481179+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the teacher trained with skeletal keypoints produces latent representations that are statistically indistinguishable from a teacher trained without them, then the entire distillation pipeline is adding architectural complexity without transferring pose-specific social information. The ablation results already approach this boundary.","supporting_citations":[],"review_version":1}