{"id":"aecac6c4-d159-486b-b000-39829dba4ed2","arxiv_id":"2506.05719","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A unified single-stage network jointly predicting semantic labels, centroid offsets, and Normalized Part Coordinate Space maps estimates category-level 6D part poses and sizes for articulated objects in real time.","lead":"YOEO is a single-stage network that estimates the 6D poses and sizes of articulated object parts from a single depth point cloud, predicting part semantics, instance centers, and canonical coordinates in one pass. The paper reports 200Hz inference and real-robot manipulation of unseen objects, with better rotation accuracy but worse translation accuracy than the two-stage GAPartNet baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-system claim rests on an unvalidated Grounding-DINO dependency: accuracy and 200 Hz are reported on object point clouds, not on the segmentation output that actually feeds the network, so end-to-end performance is unknown.","rationale":"The paper plausibly shows that jointly training semantic, centroid-offset, and NPCS heads improves part segmentation and pose metrics over individually trained heads (Re 19.6 to 9.0), and the 1.9M-parameter network plus Umeyama registration yields respectable results on GAPartNet. The real-robot experiments, though small (30 trials), demonstrate the pipeline can drive a Kinova arm and beat the baseline overall (22/30 vs 20/30), which is genuine supporting evidence. My concern is not that the network is internally wrong but that the paper's system-level central claim—single-stage, real-time, synthetic-to-real manipulation—depends on an upstream Grounding-DINO segmentation step that is never evaluated. All accuracy and speed numbers appear to be computed from already-segmented object point clouds, so detection failures are invisible in the headline metrics. I also note secondary issues: translation error is 4.6x worse than GAPartNet (0.11 vs 0.024 cm) despite the 'significantly improved pose accuracy' claim, and mIoU is a segmentation metric rather than a direct pose-accuracy metric; these should be clarified in a revision. None of this warrants rejection; the method is reproducible in principle and the conditional verdict should stand until end-to-end evidence is provided.","tokens_in":10665,"tokens_out":7056,"duration_ms":80055,"concrete_test":"Run the full published pipeline end-to-end on a held-out set of RGB-D scenes (e.g., 50 tabletop scenes with articulated objects) using Grounding-DINO with the intended text prompt; report object detection recall, pose errors on detected objects, and total latency including detection, clustering, and registration. Compare against Table I. If detection recall is below, say, 90% or end-to-end rate is far below 200 Hz, the system-level claims need to be scoped to the network only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III states that 'the input point cloud, which is obtained from the output of the segmentation model. Here, we use Grounding-DINO [40]', yet no experiment measures the impact of this upstream detector. Table I's speeds are 'calculated from feeding object point clouds to get part poses', which can be read as bypassing Grounding-DINO, and the GAPart evaluation likely uses dataset-provided object crops rather than detector outputs. If Grounding-DINO misses the object, returns a partial crop, or needs a wrong text prompt, the downstream pose estimate is undefined or degraded; the paper's own Limitations section admits small and metallic objects produce poor point clouds, exactly the regime where detection and depth are unreliable. The 200 Hz figure is therefore not end-to-end (no Grounding-DINO, clustering, RANSAC, or camera/robot overhead), and the real-robot 22/30 success rate cannot be attributed specifically to YOEO without a breakdown of detection versus pose failures. Without this evaluation, the central claim of a reliable real-time synthetic-to-real pipeline is not verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOEO, a single-stage category-level articulated-object 6D pose estimation method. A RandLA-Net backbone jointly predicts per-point semantic labels, instance centroid offsets, and Normalized Part Coordinate Space (NPCS) maps; a clustering step groups points into part instances, and Umeyama SIM(3) registration between predicted NPCS points and observed points recovers pose and size. Experiments on the GAPartNet dataset compare against PointGroup- and AutoGPart-based baselines and GAPartNet, and a real-robot study on a Kinova arm reports manipulation success rates. The paper claims faster inference (200 Hz), fewer parameters (1.9 M), and improved pose accuracy, particularly in mIoU.","tokens_in":10877,"tokens_out":2821,"duration_ms":28963,"significance":"If the claims are substantiated, YOEO would be a useful step toward practical real-time perception for articulated-object manipulation: a single lightweight network that simultaneously outputs segmentation and NPCS maps is architecturally simpler than two-stage pipelines, and the ablation in Table II provides concrete evidence that joint training of the three heads improves pose accuracy relative to individually trained heads. The real-robot deployment is also a valuable proof-of-concept. However, the central accuracy claim is currently overstated relative to the paper's own Table I, and the end-to-end real-time claim depends on an unvalidated Grounding-DINO upstream component. The core idea is defensible, but several load-bearing claims need revision or additional experiments.","major_comments":[{"comment":"The claim that YOEO achieves 'significantly improved pose accuracy compared to the previous state-of-the-art method, GAPartNet' is contradicted by the translation error in Table I: Te is 0.11 cm for YOEO versus 0.024 cm for GAPartNet, a 4.6x degradation, while rotation error improves only marginally (9.0 vs 9.9 degrees) and A5/A10 improve by only 2.1 and 1.3 percentage points. The large mIoU gain (57.6 vs 51.2) is a segmentation metric, not a pose accuracy metric. The authors should either explain why the translation error is worse (e.g., a quirk of the normalized scale or of the clustering step) or substantially qualify the headline accuracy claim.","section":"Table I and Section IV-A"},{"comment":"The input point cloud is stated to be 'obtained from the output of the segmentation model. Here, we use Grounding-DINO,' but no experiment measures the impact of this upstream detector. If Grounding-DINO misses the object, returns a partial crop, or receives an unmatched text prompt, the pose estimate is undefined or degraded; the paper's own Limitations section concedes poor behavior on small and metallic objects, which is precisely the regime where detection and depth are unreliable. The reported 200 Hz speed is 'calculated from feeding object point clouds to get part poses' (Section IV), so it excludes Grounding-DINO, clustering, RANSAC, and camera/robot overhead and is not an end-to-end rate. The authors should either evaluate the full pipeline including Grounding-DINO or clearly restate the claimed speed as network-only.","section":"Section III (Architecture Overview) and Section IV (Evaluation)"},{"comment":"The real-robot results are too weak to support the utility claim as stated. YOEO achieves 22/30 total successes versus 20/30 for GAPartNet, but on the drawer task YOEO is worse (5/10 vs 7/10). With 10 trials per task, these differences are not statistically meaningful, and the paper does not report a breakdown of failures into upstream detection failures, pose estimation failures, or execution failures. Without such a breakdown, the 22/30 rate cannot be attributed specifically to YOEO's pose estimation. The authors should add a failure analysis and, ideally, more trials.","section":"Section IV-C, Table III"},{"comment":"The ablation description is underspecified. The paper says each head is 'trained individually by freezing the other two' and then the individually trained heads are combined, but it does not state which backbone/feature-extraction parameters are frozen, whether each individually trained head uses the same trained backbone, or how the combined model is assembled at inference. Since the joint-training benefit is a central contribution, this procedure must be described precisely enough to be reproducible. The authors should clarify the training protocol and, if possible, report the variance across multiple training runs.","section":"Section IV-B, Table II"}],"minor_comments":[{"comment":"The baseline name is inconsistently spelled 'GAPartNet' and 'GAParNet' in the abstract, introduction, and Table I captions; please use 'GAPartNet' consistently.","section":"Throughout"},{"comment":"The focal loss definition is unclear: 'qi = ci · li' mixes confidence with a one-hot label; please define qi explicitly as the predicted probability for the ground-truth class, or write the loss in the standard form.","section":"Section III-A, Eq. (1)"},{"comment":"The text cites reference [44] for both RANSAC and Umeyama, but reference [44] is the RANSAC paper while Umeyama is reference [16]; the citations should be corrected.","section":"Section III-C"},{"comment":"The metric notation 'A5=5◦5cm' and 'A10=10◦10cm' is missing multiplication signs and spaces; please format as '5°, 5 cm' for readability.","section":"Table I and Table II"},{"comment":"The paper states that object scales in the dataset are normalized to a range of 0 to 1 cm; this unusual normalization should be explained, since it directly affects the interpretation of Te and Se values.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a coherent single-stage architecture and an ablation that supports joint training, but the authors' own Table I undermines the headline accuracy claim, and the end-to-end real-time claim rests on an unevaluated Grounding-DINO dependency. These are fixable with revised claims and additional experiments, so I recommend major revision rather than rejection. I would also encourage the editor to ask the authors to make the training/evaluation protocol more reproducible, including the exact split of the GAPartNet dataset used for the comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the synthesis: one shared network predicts semantic labels, centroid offsets, and NPCS maps, and the ablation shows co-training is what makes it work. Rotation error drops from 19.6 to 9.0 degrees when you train the heads together instead of separately. That is a real, non-obvious result, and it is the best evidence in the paper. The architecture is also light: 1.9M parameters and fast network inference, both sensible given the RandLA-Net backbone and shared MLPs.\n\nI agree with the reader that the paper deserves a serious referee, but the soft spots are real and should be addressed. The most obvious is Table I: the paper claims \"significantly improved pose accuracy\" while reporting a translation error of 0.11 cm, which is 4.6 times worse than GAPartNet's 0.024 cm. Rotation improves only marginally (9.0 vs 9.9 degrees). That gap is never discussed, and it undercuts the headline. The mIoU gain and A5/A10 improvements are legitimate, but the translation regression should have been acknowledged.\n\nThe second soft spot is the 200Hz claim. The paper states speeds are calculated from feeding object point clouds directly, which means the Grounding-DINO segmentation step is not included. The stress-test note is right: the real-system claim depends on an upstream detector whose failures are not evaluated. The paper's own limitations admit poor performance on small and metallic objects, exactly the regime where detection and depth are least reliable. So the end-to-end latency and robustness are unknown. This is not a fatal flaw, but it means the \"real-time synthetic-to-real pipeline\" claim is only partially verified. The robot experiments show 22/30 vs GAPartNet's 20/30, which is a modest difference and not broken down by failure cause.\n\nThe missing code, error bars, and training details are minor but standard issues for a systems paper; they are addressable.\n\nOverall, the core idea is novel, the ablation is informative, and the literature coverage is fine. The paper overstates its accuracy advantage and leaves the detector dependency unexamined, but those are fixable. I would send it to peer review and ask the authors to explain the translation error, report end-to-end latency including detection and clustering, and add variance estimates.","headline":"A genuinely new single-stage architecture for articulated part pose with a strong joint-training ablation, but the headline accuracy claim overstates a worse translation error and the real-time claim skips the upstream detector.","tokens_in":11437,"tokens_out":1718,"would_cite":true,"duration_ms":18507,"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":"A single 200 Hz pass estimates articulated-part 6D poses.","keywords":["6D pose estimation","articulated objects","category-level pose estimation","robotic grasping","point cloud instance segmentation","NPCS","single-stage network","real-time manipulation"],"falsifier":"Run YOEO on a fixed set of GAPart test objects using two input sources: object crops from Grounding-DINO text prompts and ground-truth object masks from the dataset. If rotation error and mIoU stay comparable across the two, the upstream dependency is weak; if accuracy drops sharply for small or metallic instances or under occlusion, the dependency is confirmed. A complementary test is to artificially decimate or partially mask the input point cloud and record at what coverage the rotation error exceeds 15 degrees.","tokens_in":1675,"feed_emoji":"🤖","tokens_out":2092,"duration_ms":73512,"temperature":0.7,"pith_summary":"This paper claims that category-level pose estimation for articulated objects does not need a two-stage pipeline that first segments part instances and then estimates Normalized Part Coordinate Space (NPCS) representations. YOEO is a single-stage point-cloud network that jointly predicts per-point semantic labels, offsets to part-instance centroids, and NPCS coordinates, then clusters points by their voted centroids and aligns each NPCS region to the observed cloud with Umeyama registration to recover pose and size. On the GAPart dataset, the authors report improved rotation error (9.0 degrees versus 9.9 degrees) and 3D mean intersection-over-union (57.6 percent versus 51.2 percent) over the two-stage GAPartNet baseline, with only 1.9 million parameters and inference at 200 Hz. They also deploy the synthetically trained model on a real robot to manipulate unseen articulated objects, which demonstrates that a single forward pass can support real-time robotic perception.","feed_headline":"One-stage network beats two-stage part pose estimation at 200 Hz","feed_subtitle":"Jointly learning semantic labels, centroid votes, and NPCS maps cuts rotation error from 19.6° to 9.0°.","key_machinery":"The load-bearing object is the unified one-stage architecture: a RandLA-Net encoder-decoder whose features feed three parallel shared-MLP heads—semantic labels trained with focal loss, per-point centroid offsets trained with L1 loss, and NPCS coordinates trained as 100-bin classification per axis. The mechanism that separates instances is centroid voting: every point predicts a vector to its part's centroid, so points of the same instance agree and different instances of the same semantic class separate in offset space; a clustering step groups them, and RANSAC plus Umeyama least-squares registration recovers the similarity transform from the predicted NPCS region to the observed point cloud.","core_discovery":"The central claim is that a single forward pass of one unified network can replace the cascade of part segmentation followed by per-part NPCS estimation, and that co-training the three prediction heads—semantic labeling, centroid-offset voting, and NPCS regression—is what produces the accuracy gain rather than a more complex architecture. The paper's evidence is the ablation study: parallel training reduces rotation error from 19.6 degrees to 9.0 degrees, reduces translation and scale errors from 0.14 cm to 0.11 cm and from 0.041 cm to 0.036 cm, and lifts 3D mIoU from 52.3 percent to 57.6 percent compared to individually trained heads. The full model reaches 200 Hz inference with 1.9 million parameters, and a physical Kinova robot completes 22 of 30 manipulation trials on objects not seen during training.","pith_inferences":["The large jump between individual and parallel rotation error suggests the auxiliary heads act as a learned regularizer for the NPCS head; one could test this by adding the auxiliary losses incrementally to a pure NPCS network and measuring the gain at each step.","The pipeline's reliance on Grounding-DINO for the input crop means the whole system's accuracy is upper-bounded by the detector's recall; an evaluation that reports pose error conditioned on detector success versus failure would expose how much of the sim-to-real result depends on that upstream stage.","Since the paper lists small and metallic objects as failure cases due to depth quality, fusing RGB cues from the same crop is a natural extension, and its value could be quantified by measuring rotation error on those object classes with and without RGB fusion."],"forward_implications":["A one-stage formulation removes the error accumulation that comes from cascaded segmentation-then-pose pipelines.","Co-training three auxiliary heads is a practical recipe for improving NPCS regression accuracy without adding new supervision.","Because the predicted NPCS frame carries the joint or prismatic axis and predefined grasp poses, the same registered transform directly yields robot motion commands.","At 200 Hz and 1.9 million parameters, the perception cost drops enough for on-board, real-time control loops rather than offline pose readouts."],"supporting_citations":[{"why":"Provides the GAPart dataset and the two-stage GAPartNet baseline that YOEO is compared against.","marker":"[14]"},{"why":"Supplies the RandLA-Net backbone used for point-cloud feature extraction in the unified encoder-decoder.","marker":"[15]"},{"why":"Gives the Umeyama least-squares registration used to recover the SIM(3) pose and size from the NPCS region.","marker":"[16]"},{"why":"Introduces Normalized Object Coordinate Space, the category-level canonical-space representation that NPCS adapts for articulated parts.","marker":"[37]"},{"why":"Grounding-DINO is the vision-language model that produces the object point-cloud crop that is the network's input.","marker":"[40]"},{"why":"Represents the prior articulated-object pose method whose two-stage formulation this paper aims to replace.","marker":"[11]"},{"why":"Supplies the PointGroup-based baseline (PG) adapted to part pose estimation and compared in the results table.","marker":"[45]"},{"why":"Supplies the AutoGPart-based baseline (AGP) used for comparison in the results table.","marker":"[46]"}],"fun_headline_variants":["One-stage pose estimation hits 200 Hz for articulated objects","Single unified network beats two-stage for pose at 200 Hz","YOEO: one pass to 6D pose, 200 Hz, 9° error","Single-stage method: 200 Hz and half the rotation error","Unified network does pose and size in one forward pass"],"cache_read_input_tokens":13568,"weakest_assumption_plain":"The load-bearing premise is that the point cloud delivered to the network is a clean, complete crop of the target object from Grounding-DINO, and that the depth sensor provides usable geometry for the part; failures of that upstream detector or depth acquisition are not evaluated, and the paper's own limitations concede poor performance on small and metallic objects.","fun_headline_variants_meta":{"raw":{"variants":["One-stage pose estimation hits 200 Hz for articulated objects","Single unified network beats two-stage for pose at 200 Hz","YOEO: one pass to 6D pose, 200 Hz, 9° error","Single-stage method: 200 Hz and half the rotation error","Unified network does pose and size in one forward pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1731,"prompt_tokens":986,"completion_tokens":745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":654}},"tokens_in":602,"tokens_out":745,"duration_ms":8196,"temperature":1.0,"reasoning_tokens":654,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:13:18.718830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run YOEO on a fixed set of GAPart test objects using two input sources: object crops from Grounding-DINO text prompts and ground-truth object masks from the dataset. If rotation error and mIoU stay comparable across the two, the upstream dependency is weak; if accuracy drops sharply for small or metallic instances or under occlusion, the dependency is confirmed. A complementary test is to artificially decimate or partially mask the input point cloud and record at what coverage the rotation error exceeds 15 degrees.","supporting_citations":[{"cited_title":"Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts,","cited_arxiv_id":null,"evidence_quote":"Provides the GAPart dataset and the two-stage GAPartNet baseline that YOEO is compared against."},{"cited_title":"Randla-net: Efficient semantic segmentation of large- scale point clouds,","cited_arxiv_id":null,"evidence_quote":"Supplies the RandLA-Net backbone used for point-cloud feature extraction in the unified encoder-decoder."},{"cited_title":"Least-squares estimation of transformation parameters between two point patterns,","cited_arxiv_id":null,"evidence_quote":"Gives the Umeyama least-squares registration used to recover the SIM(3) pose and size from the NPCS region."},{"cited_title":"Category-level articulated object pose estimation,","cited_arxiv_id":null,"evidence_quote":"Represents the prior articulated-object pose method whose two-stage formulation this paper aims to replace."},{"cited_title":"Pointgroup: Dual-set point grouping for 3d instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the PointGroup-based baseline (PG) adapted to part pose estimation and compared in the results table."},{"cited_title":"Autogpart: Intermediate supervision search for generalizable 3d part segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the AutoGPart-based baseline (AGP) used for comparison in the results table."}],"review_version":1}