{"id":"0372c84c-731e-438f-b0e1-721636e5bee6","arxiv_id":"2507.22522","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ACTIVE is a new RGB+LiDAR action recognition dataset from moving robotic viewpoints at 3-50m, and ACTIVE-PC improves point cloud action recognition on it to 60.10% Top-1 accuracy.","lead":"This paper introduces ACTIVE, a large multi-modal video dataset of people performing actions while a moving camera and LiDAR record from 3 to 50 meters away. It also presents ACTIVE-PC, a point cloud method that raises action recognition accuracy on this hard benchmark from 46.76% to 60.10%.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Point-cloud accuracy in Table 2 is conditional on an unquantified TransFusion person-detection stage; without distance-stratified detection/recognition results, the 60.10% may not measure ACTIVE-PC's action understanding.","rationale":"The paper's central claim rests on the benchmark value of ACTIVE and the effectiveness of ACTIVE-PC. The benchmark value is demonstrated mainly by the point cloud numbers in Tables 2 and 5; if those numbers are confounded, both the method and the dataset evaluation lose support. The unvalidated TransFusion preprocessing is the least secure link because it sits before every point-cloud experiment and the paper provides no evidence that it works in the specific regime (3-50m, HESAI AT128P, day/night, moving platform) that defines ACTIVE. The paper's own text in Sec. 5.1 is the only description of this stage, and it omits detection metrics, distance-stratified analysis, and even how detections are converted to 768-point clips. The other weaknesses the reader notes (single runs, no attribute results, no explicit dataset link) would lower confidence but do not by themselves threaten the central comparison; even a false 'first large-scale N-HRI dataset' claim (e.g., in light of HOMAGE in Table 1) would affect novelty more than the validity of Table 2. My proposed test is deliberately minimal: hold detector quality fixed or compare against ground-truth crops and stratify by distance. If detection is the bottleneck, the paper's long-range action-recognition claim becomes conditional; if accuracy per distance holds under ground-truth crops, the concern is resolved. Therefore the reader's CONDITIONAL verdict is the right one.","tokens_in":12268,"tokens_out":9343,"duration_ms":108633,"concrete_test":"Annotate a stratified subset of ACTIVE point-cloud clips with ground-truth person point sets (or bounding boxes) in distance bins: 3-10m, 10-25m, 25-50m. Then run ACTIVE-PC and the strongest baseline twice: once with the paper's TransFusion detections and once with ground-truth person crops, using the same 768-point resampling and 12-frame protocol. If (i) TransFusion AP/recall is low in the 25-50m bin, or (ii) the ACTIVE-PC vs PST-Transformer gap shrinks substantially when ground-truth crops replace detections, then Table 2 mainly measures detector recall and the long-range claim is conditional at best.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative anchor of the central claim is Table 2: ACTIVE-PC reaches 60.10% Top-1 on ACTIVE point clouds, 13.34 points above PST-Transformer. But every number in this table is downstream of a preprocessing stage that the paper does not validate. Section 5.1 states: 'In the point cloud modality, we first apply the TransFusion [1] method to detect human subjects within the scene. The point cloud sequence is then segmented into fixed-length clips, with each frame sampled to 768 points.' No detection recall, precision, distance-stratified analysis, or adaptation details for TransFusion are reported. TransFusion is a camera-LiDAR fusion 3D detector developed for driving scenes; ACTIVE uses a HESAI AT128P LiDAR over 3-50m with 0.1°/0.2° resolution, so at 50m a 1.8m tall person spans a very small number of beams/points. If person recall drops at long range, the subsequent resampling to 768 points will tend to sample background, and Table 2 becomes a joint detection+action-recognition score rather than a measure of ACTIVE-PC's ability to perceive human actions. The same gap weakens the 'long-distance' motivation: the paper never reports accuracy as a function of distance. Thus the claim 'accurately perceives human actions at long distances' is not evidenced by the aggregate 60.10%.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ACTIVE, a large-scale multimodal (RGB + LiDAR point cloud) action recognition dataset for Natural Human-Robot Interaction, comprising 46,868 annotated instances, 80 subjects, 30 composite action categories, six environments, and capture distances from 3 to 50 meters from a mobile platform. The authors also propose ACTIVE-PC, a point cloud action recognition method built on Multilevel Neighborhood Sampling (MNS), Layered Recognizers (LR), and Elastic Ellipse Query (EEQ). They report 60.10% Top-1 accuracy on ACTIVE point clouds, outperforming PST-Transformer by 13.34 points, and 91.7% CS / 96.8% CV on NTU RGB+D. An RGB variant, ACTIVE-RGB, is also evaluated on ACTIVE, reaching 57.56% Top-1 accuracy.","tokens_in":12559,"tokens_out":4370,"duration_ms":52778,"significance":"If the benchmark and method claims hold, ACTIVE would fill a real gap: existing action recognition datasets do not combine long capture distances, platform motion, multimodal sensing, and N-HRI-oriented composite actions at this scale. ACTIVE-PC's gains on ACTIVE and its consistent performance on the external NTU RGB+D benchmark suggest the implementation is not broken and the architecture is worth studying. The authors also provide a code release, which supports reproducibility. However, the central quantitative claims currently rest on an unquantified person-detection preprocessing stage and single-run evaluations, so the evidence as presented is not yet sufficient to establish the benchmark's validity or the method's state-of-the-art status.","major_comments":[{"comment":"The point cloud evaluation pipeline begins with TransFusion person detection, as stated in §5.1: \"we first apply the TransFusion [1] method to detect human subjects within the scene.\" No detection accuracy, precision/recall, or distance-stratified analysis is reported. Because ACTIVE spans 3 to 50 meters and each frame is resampled to 768 points, distance-dependent detector failures would inject background points into the input, making the 60.10% accuracy in Table 2 a joint detection-plus-recognition score rather than a measure of ACTIVE-PC's action understanding. Please report TransFusion detection performance stratified by distance, provide failure-case analysis, or evaluate using ground-truth human crops, and re-present Table 2 accordingly.","section":"§5.1, Table 2"},{"comment":"All accuracy numbers in Tables 2, 4, 5, and 6 are single-run point estimates on a single cross-subject split (53 training / 27 test subjects). Without error bars or significance tests, the 13.34-point improvement over PST-Transformer and the ablation differences in Table 5 could be influenced by seed variability. Please report mean and standard deviation over at least three independent runs and, if possible, additional subject splits or a statistical comparison.","section":"§5.1, Tables 2 and 4"},{"comment":"The dataset characterization is incomplete for a benchmark paper. Section 3 reports totals (46,868 instances, 80 subjects, 30 classes, 6 sites) but does not provide per-class instance counts, distance distribution, day/night and environment breakdowns, or an annotation protocol with inter-annotator agreement. These details are necessary to assess class balance, task difficulty, and annotation reliability. Additionally, the claim of being \"the first large-scale action recognition dataset specifically focused on N-HRI\" should be supported by a more explicit comparison with the closest existing datasets, such as HOMAGE and SMG, rather than only by the summary row in Table 1.","section":"§3, Table 1"},{"comment":"The abstract claims that ACTIVE-PC achieves \"precise decoupling of kinematic interference from human actions,\" but the fusion mechanism in Eq. (1) is a simple average of Yhuman and Ykinematic. There is no explicit mechanism that subtracts or orthogonalizes kinematic components from action features, and no experiment demonstrates that the Kinematic Interpreter actually isolates platform motion. Please either soften the decoupling claim or provide evidence, such as kinematic label prediction accuracy, feature visualizations, or a controlled experiment varying platform motion.","section":"§4.4, Eq. (1)"}],"minor_comments":[{"comment":"The figure caption or figure itself contains the leftover artifact \"1.png2.png3.png4.png\"; this should be removed.","section":"Figure 3"},{"comment":"Equation (1) has an unmatched parenthesis: it reads \"Y = (Yhuman + Ykinematic)) / 2\" with one extra closing parenthesis.","section":"Eq. (1)"},{"comment":"Reference [18] is cited as \"I Loshchilov. Decoupled weight decay regularization\" without a venue or year; please provide the full citation.","section":"References"},{"comment":"The text states that \"Transformer-based models outperform CNN-based models,\" but Table 2 does not include any CNN-based baseline; either add a CNN baseline or revise the sentence to match the reported comparisons.","section":"§5.2"},{"comment":"The t-SNE visualization covers only the 11 Base Motions categories, not all 30 action classes; the caption should state this limitation so that the claim about separating similar actions is not overgeneralized.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ACTIVE is a genuinely new resource: 46,868 instances, 80 subjects, RGB and LiDAR, 3–50 m range, mobile platform, six sites, day/night. The composite action labels and the explicit focus on perception-centric robotic views fill a gap that NTU and the micro-gesture datasets leave open. That alone makes the paper worth taking seriously. ACTIVE-PC is less novel but still sensible: MNS and Layered Recognizers are engineering choices, and EEQ's learned anisotropic ellipse is a modest but reasonable idea. The ablations show each piece helps, the NTU results are consistent with prior point-cloud methods, and Table 6 shows EEQ does reduce sensitivity to the neighborhood radius. The code link is a plus.\n\nThe reader's stress-test concern is real. Section 5.1 says the point cloud pipeline's first step is TransFusion person detection, but the paper gives no detection recall, precision, or distance-stratified quality. At 50 m a 1.8 m person is a handful of LiDAR points; if detection drops at range, the 60.10% number is a joint detection-plus-recognition score, not an action-recognition score. The paper also never reports accuracy versus distance, so the long-range motivation is not actually evidenced. This is a load-bearing gap, not a nit. A related soft spot: attribute recognition is introduced as a benchmark task in Sec. 3.2 but never evaluated anywhere. And while results are compared across many baselines, there are no error bars; on a brand-new benchmark, single-run accuracies should be treated with caution.\n\nTo be fair, these are fixable. Because every baseline in Table 2 runs through the same preprocessing, the method-versus-baseline comparison is probably still meaningful. The dataset itself is valuable whether or not ACTIVE-PC's absolute accuracy holds up. What is missing is the supporting evidence for the specific claim that the method \"accurately perceives human actions at long distances.\"\n\nI would accept this for peer review. A serious referee should ask for distance-stratified detection and recognition results, ground-truth-box experiments, attribute-recognition baselines, and a clear dataset release statement. With those, it could be a useful benchmark paper. Without them, the method claims run ahead of the evidence.","headline":"A genuinely useful new dataset for HRI action recognition that deserves review, but the point-cloud headline numbers need distance-stratified detection analysis before they support the long-range claims.","tokens_in":13134,"tokens_out":2550,"would_cite":true,"duration_ms":30652,"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":"The paper introduces ACTIVE, a 46,868-instance multimodal benchmark of human actions filmed from moving, perception-centric robot views at 3–50 m, and claims ACTIVE-PC recognizes them with 60.10% top-1 point cloud accuracy, 13.34 points…","keywords":["action recognition","human-robot interaction","robotic view","point cloud video","LiDAR","RGB video","dataset benchmark","multi-modal perception"],"falsifier":"Measure TransFusion's person-detection recall on ACTIVE in distance bins (for example 3–10 m, 10–20 m, and 20–50 m) and compare ACTIVE-PC's top-1 accuracy on correctly detected frames against frames where the detector fails; low long-range recall would show that the reported 60.10% partly measures detector performance rather than action recognition.","tokens_in":12057,"feed_emoji":"🤖","tokens_out":7179,"duration_ms":68624,"temperature":0.7,"pith_summary":"The paper introduces ACTIVE, a large-scale benchmark that fixes a blind spot in human action recognition: most datasets capture actions at fixed, close range from static cameras, whereas real service robots see people from a moving platform at distances of 3 to 50 meters. ACTIVE records 30 composite action classes from 80 subjects in 46,868 annotated instances, with synchronized RGB and LiDAR point cloud views across indoor/outdoor and day/night scenes, including platform motion and vibration. The authors claim this is the first large-scale action recognition dataset built specifically for Natural Human-Robot Interaction, and they propose ACTIVE-PC to recognize actions on it. On the point cloud modality, ACTIVE-PC reaches 60.10% top-1 accuracy, beating PST-Transformer's 46.76% by 13.34 percentage points, and its RGB counterpart outperforms general video models on the same benchmark. If the claims hold, the field gains a shared testbed for the conditions robots actually face: long distances, ego-motion, and subtle human actions.","feed_headline":"Robot-view action benchmark lifts top-1 accuracy to 60%","feed_subtitle":"ACTIVE-PC beats prior point-cloud methods by 13.3 points on actions filmed from a moving robot at 3–50 meters.","key_machinery":"The machinery is a three-part point cloud video architecture. Multilevel Neighborhood Sampling (MNS) replaces global-only farthest point sampling with hierarchical local-neighborhood sampling, preserving global coverage while retaining fine local detail at long range. Layered Recognizers (LR) split features by density: a Kinematic Interpreter pools sparse layers with global average pooling to model robot-platform and human-robot motion, while an Action Recognizer pools the dense layer with global max pooling to keep subtle body motions, and the two score streams are averaged. Elastic Ellipse Query (EEQ) replaces the spherical Ball Query with a learnable axis-specific metric $d = \\sqrt{\\alpha(x_n-x_q)^2 + \\beta(y_n-y_q)^2 + \\gamma(z_n-z_q)^2}$, so planar shifts from robot motion are compressed while vertical posture changes stay sensitive; on ACTIVE the learned scales are $(\\alpha,\\beta,\\gamma) = (3.5632, 3.6789, 2.8038)$. These components carry the claimed 13.34-point gain, and the ablation table shows removing any of them lowers accuracy.","core_discovery":"The paper's central claim is that Natural Human-Robot Interaction requires its own action recognition benchmark because conventional datasets assume fixed, close-range, static-camera views. ACTIVE fills that gap with 30 composite action classes, 80 subjects, and 46,868 video instances in RGB and LiDAR point cloud modalities, captured at 3–50 meters across six environments, day and night, with a moving and vibrating capture platform. On this benchmark, ACTIVE-PC reaches 60.10% top-1 accuracy in the point cloud modality, surpassing the strongest baseline (PST-Transformer at 46.76%) by 13.34 percentage points; on NTU RGB+D it remains competitive (91.7% cross-subject, 96.8% cross-view), and the RGB variant ACTIVE-RGB outperforms InternVideo2 by 2.68% on ACTIVE. The intended reading is that the proposed components, Multilevel Neighborhood Sampling, Layered Recognizers, and Elastic Ellipse Query, are what make the gain possible rather than a larger backbone or more data.","pith_inferences":["Extension: the learned anisotropic metric of EEQ could transfer to other egocentric or robot-view point cloud tasks, such as pedestrian intent prediction or mobile-robot navigation safety, where ego-motion dominates the point distribution.","Extension: because the dataset includes both action labels and kinematic disturbance from platform motion, future work could report a separate robustness score that isolates ego-motion invariance from raw accuracy, which the current benchmark does not define.","Extension: a testable follow-up is to train the detector and ACTIVE-PC jointly; if long-range detection is the bottleneck the paper does not quantify, end-to-end training could improve the 60.10% anchor more than any action-recognition module change.","Extension: the 3–50 m distance range invites distance-stratified evaluation, for example reporting accuracy separately for short, medium, and long ranges; the paper reports only aggregate top-1 accuracy."],"forward_implications":["General point cloud action methods transfer poorly to robotic views: the best prior baseline reaches only 46.76% on ACTIVE, so the dataset provides a way to measure the gap between clean, close-range action recognition and N-HRI conditions.","The decoupling of kinematic state from human action appears to matter: removing Layered Recognizers or EEQ each costs roughly five accuracy points on ACTIVE.","EEQ makes performance less sensitive to the neighborhood radius hyperparameter, with accuracy variance dropping from 0.456 to 0.100 across radius settings.","The RGB modality also benefits from the robotic-view design: ACTIVE-RGB (57.56% at $12\\times224$) outperforms InternVideo2 (54.88%) on ACTIVE.","With 46,868 instances carrying action and attribute labels, the benchmark supports joint action and attribute recognition, and pose annotations are planned as a later release."],"supporting_citations":[{"why":"Supplies the TransFusion detector that localizes human subjects in the point cloud before frames are resampled to 768 points, so every point cloud accuracy number depends on this detector.","marker":"[1]"},{"why":"P4Transformer is a point cloud video baseline (43.88%) that ACTIVE-PC must beat and whose transformer architecture the point tube builds on.","marker":"[6]"},{"why":"PST-Transformer is the strongest point cloud baseline (46.76%) that ACTIVE-PC surpasses by 13.34 percentage points in Table 2.","marker":"[9]"},{"why":"3DinAction is the most recent general point cloud action method (49.19%) used as a competitor to show ACTIVE-PC's advantage on the new benchmark.","marker":"[2]"},{"why":"NTU RGB+D is the standard large-scale action dataset used to verify that ACTIVE-PC retains competitive accuracy (91.7% cross-subject) on general action recognition.","marker":"[21]"},{"why":"YOLOv10 is the object detector that locates human bodies in RGB frames before clip-level recognition in the ACTIVE-RGB pipeline.","marker":"[26]"}],"fun_headline_variants":["Robot-view action benchmark: 46,868 clips, 3-50m","ACTIVE-PC hits 60% top-1 on robot-view actions","Moving robot action recognition gets its own dataset","Robot-view LiDAR action model beats PST by 13.3%","New robot-view dataset and model for natural HRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The point cloud results assume that the TransFusion detector reliably finds every human subject in every frame across the full 3–50 m range before frames are resampled to 768 points (Section 5.1), and the paper does not report detection accuracy or failure cases by distance.","fun_headline_variants_meta":{"raw":{"variants":["Robot-view action benchmark: 46,868 clips, 3-50m","ACTIVE-PC hits 60% top-1 on robot-view actions","Moving robot action recognition gets its own dataset","Robot-view LiDAR action model beats PST by 13.3%","New robot-view dataset and model for natural HRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2182,"prompt_tokens":1032,"completion_tokens":1150,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1061}},"tokens_in":648,"tokens_out":1150,"duration_ms":11768,"temperature":1.0,"reasoning_tokens":1061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:34:53.979084+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure TransFusion's person-detection recall on ACTIVE in distance bins (for example 3–10 m, 10–20 m, and 20–50 m) and compare ACTIVE-PC's top-1 accuracy on correctly detected frames against frames where the detector fails; low long-range recall would show that the reported 60.10% partly measures detector performance rather than action recognition.","supporting_citations":[{"cited_title":"Transfusion: Robust lidar-camera fusion for 3d object detection with transform- ers","cited_arxiv_id":null,"evidence_quote":"Supplies the TransFusion detector that localizes human subjects in the point cloud before frames are resampled to 768 points, so every point cloud accuracy number depends on this detector."},{"cited_title":"Point 4d trans- former networks for spatio-temporal modeling in point cloud videos","cited_arxiv_id":null,"evidence_quote":"P4Transformer is a point cloud video baseline (43.88%) that ACTIVE-PC must beat and whose transformer architecture the point tube builds on."},{"cited_title":"Point spatio- temporal transformer networks for point cloud video mod- eling","cited_arxiv_id":null,"evidence_quote":"PST-Transformer is the strongest point cloud baseline (46.76%) that ACTIVE-PC surpasses by 13.34 percentage points in Table 2."},{"cited_title":"3di- naction: Understanding human actions in 3d point clouds","cited_arxiv_id":null,"evidence_quote":"3DinAction is the most recent general point cloud action method (49.19%) used as a competitor to show ACTIVE-PC's advantage on the new benchmark."},{"cited_title":"Ntu rgb+ d: A large scale dataset for 3d human activity anal- ysis","cited_arxiv_id":null,"evidence_quote":"NTU RGB+D is the standard large-scale action dataset used to verify that ACTIVE-PC retains competitive accuracy (91.7% cross-subject) on general action recognition."}],"review_version":1}