{"id":"ad53149d-7a23-46bf-b17d-06e6fb170b1f","arxiv_id":"2412.01083","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A sim-to-real pipeline decomposes robotic hanger insertion into two learned action primitives and achieves 75% success on eight unseen real garments.","lead":"Roboticists built a cloth simulator and used it to train a two-armed robot to slip a hanger through the necklines of garments. The learned policy transferred to eight unseen real garments with a 75% success rate, without any real-world fine-tuning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 75% real-world success rate is anchored by a manually tuned simulation success metric and only 16 real trials; until the ray-cast coverage criterion is calibrated against true insertion, the headline number remains conditional.","rationale":"The paper makes a genuinely interesting contribution: a modular action-primitive decomposition, a purpose-built cloth simulator, a scalable data-collection pipeline, and a real-robot demonstration with baselines and ablations. My stress-test did not find a fatal internal inconsistency. The central risk is empirical calibration: the simulation success metric (Sec. III-D) is a geometric proxy with manually chosen thresholds, used both for training labels and for the headline simulation number, but it has never been compared with real-world ground truth. The real-world evaluation is the only calibration point, and it is only 16 trials per condition, yielding a wide confidence interval and no pre-registered success-check protocol. The explicit limitations in Sec. V (attachment-style gripper modeling; kinematic hanger motion in Sec. III-F) are plausible sources of sim-only exploitation, and the 16-trial real experiment is too weak to rule them out. I therefore keep the reader's CONDITIONAL verdict rather than accepting the 75% figure at face value; the concrete checks above would either confirm the metric's validity or reveal the need to revise the headline. This is not a claim that the authors are wrong; it is a request for the calibration evidence that the central claim requires.","tokens_in":10953,"tokens_out":11004,"duration_ms":95377,"concrete_test":"Ask the authors to re-run the trained policies on the 24 held-out simulation garments and independently re-annotate every successful rollout by requiring that, after the final lift, (i) both hanger endpoints are visually through the neckline and (ii) the garment remains draped when the grippers release. Compare this re-annotated S to the reported 85.1%; a drop of more than 10 points would show the ray-cast thresholds are too permissive. Separately, pre-register an automated real-world success check (endpoints inside the garment in depth/RGB plus drape on release) and run at least 50 trials per condition on the same 8 garments; if the re-measured real-world S falls below about 60%, the headline 75% should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To accept the headline claim—75% on unseen real garments without fine-tuning—one must trust the simulation's definition of 'successful insertion' (Sec. III-D): an endpoint counts as inserted when at least 0.95 (left) or 0.9 (right) of rays cast from the endpoint hit the garment mesh. These thresholds are set manually and are never validated against real-world ground truth. The same criterion labels every training trajectory and computes the reported sim success rate S=85.1%, so an over-permissive threshold would make both the learned policies and the sim evaluation reward 'mesh coverage' rather than actual passage of the hanger through the neckline. The only ground-truth anchor is the real-world experiment, which is 16 trials per condition (Sec. IV-A) and gives 75% = 12/16 (95% CI roughly 51–90%). With such a small sample and no stated success-checking protocol, the real-world result cannot reliably calibrate the sim metric. The paper's own Limitations (Sec. V) acknowledge that gripper-garment contact is idealized as attachment and that hanger motion is kinematic (Sec. III-F); because these idealizations are baked into the data-generation and success-checking loop, the learned keypoints may exploit sim-only effects that the 16 real trials are too weak to detect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes RoboHanger, a dual-arm robotic system that inserts a clothes hanger into a garment through its neckline using two learned action primitives: press-and-lift for the left endpoint and drag-and-rotate for the right endpoint. Each primitive is parameterized by two 2D keypoints predicted by four U-Net value maps from depth images and object masks. Training data are generated automatically in a custom Taichi/FEM cloth simulator with 144 synthetic garments, using an initial heuristic policy followed by iterative online fine-tuning. The paper reports an overall simulation success rate of 85.1% on held-out garments and a real-world success rate of 75% on 8 unseen garments without real-data fine-tuning, outperforming IL, ACT, DP, SAC, and a fixed-policy baseline in the reported comparisons. The authors explicitly acknowledge limitations in the simulator's grasp modeling and kinematic hanger motion.","tokens_in":1517,"tokens_out":3262,"duration_ms":62769,"significance":"If the reported results hold, this is a useful contribution to deformable-object manipulation. The task decomposition into small action primitives is clean, the automatic data-collection loop in a GPU cloth simulator is scalable, and the 75% sim-to-real transfer without real fine-tuning is a concrete and falsifiable generalization claim. The use of depth and masks is well motivated, and the ablations on online data and neckline masks provide internal evidence that the design choices matter. The main open risk is whether the simulation success metric and the small real-world evaluation actually certify the headline claim; both concerns are addressable with additional calibration and trials.","major_comments":[{"comment":"The simulation success criterion is load-bearing but not calibrated to real-world ground truth. Section III-D states that an endpoint is counted as inserted when the fraction of rays hitting the garment mesh exceeds 0.95 for the left endpoint and 0.9 for the right endpoint, and that these thresholds are set manually by inspecting simulated trajectories. Because this same criterion provides training labels for Qpress, Qlift, Qdrag, and Qrotate and also produces the simulation success rates, an over-permissive threshold would reward mesh coverage rather than true insertion through the neckline. I request a validation study comparing the ray-cast coverage score against manually or externally labeled true insertion in sim and real episodes, and a sensitivity analysis around the chosen thresholds.","section":"III-D, IV-A"},{"comment":"The real-world evaluation is too small to support the strength of the claims. Section IV-A reports 16 real trials per condition, and Table III shows 75.0% versus 62.5% for the best baseline; with 16 trials, the difference is only two trials. There are no confidence intervals, no per-garment breakdown, and no stated protocol for checking real-world success. Please report exact trial counts per garment, a written success-checking protocol, confidence intervals, and preferably additional trials per condition.","section":"IV-A, Table III"},{"comment":"The idealized simulator assumptions acknowledged in the Limitations section interact directly with the training and success-checking loop. Section III-F models garment grasping as attaching the garment to the gripper and treats hanger motion as purely kinematic, so all training data, ablations, and simulated success rates are generated under these idealizations. Please provide targeted evidence that the learned keypoints do not exploit sim-only effects, for example by retraining under a more faithful grasp or contact model, or by analyzing real-world failures to attribute them to the known idealizations. At minimum, the discussion should quantify how much of the 25% real-world failure rate is plausibly due to these simplifications.","section":"III-F, V"},{"comment":"The baseline comparisons are not apples-to-apples on training data. Section IV-B states that ACT and DP are trained on about 10,000 successful trajectories from the heuristic policy, while the proposed method is trained on 18,000 trajectories that include online data collected by the learned policy. The improvements over ACT and DP in Table I could therefore reflect differences in data quantity and selection rather than the action-primitive formulation alone. Please retrain baselines on the same number of trajectories with the same offline/online mix, or include baseline numbers in the Table II data ablations.","section":"IV-B, IV-D"}],"minor_comments":[{"comment":"In the press-and-lift description, the sentence 'we then lift the garment at point plift with the other hand without moving the garment' is ambiguous; it should state which object remains stationary.","section":"III-C"},{"comment":"The notation S2 is used for the success of the right-endpoint primitive, but Table III reports S2 as 80.0% even though the overall S is 75.0% when S1 is 93.8%; please define the conditional nature explicitly.","section":"III-D"},{"comment":"The supervision in Eq. (2) and Eq. (5) uses a 0.5 threshold on the max Q value; please justify this choice or note its sensitivity.","section":"III-E"},{"comment":"The SAC baseline description says the action is first transformed into an image where each pixel value represents the inverse of the distance to that action, but the exact rendering of this action image is not fully specified; a short pseudocode or channel diagram would help reproducibility.","section":"IV-B"},{"comment":"Figure 6 labels hangers with (a.1) and (a.2) in the text of Section IV-B, but Figure 6(a) appears to use numbered hangers; please align the labels.","section":"IV-B, Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is sound, but the version I reviewed does not yet certify the headline real-world success rate. The most important fix is validating the simulation success metric against real ground truth; the real-world sample size and protocol should also be strengthened. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution is the task itself: nobody has done hanger-into-neckline insertion on flat garments with learned policies, and this paper gives a complete pipeline—two hand-crafted action primitives, four UNets predicting keypoints from depth and masks, a custom Taichi-based cloth simulator, and iterative data collection that starts from a heuristic policy and fine-tunes with network-collected data. The simulation results are strong (85.1% overall), and the real-world 75% on 8 unseen garments without real fine-tuning is a genuine sim-to-real result, not a demo with a scripted policy. The authors also deserve credit for testing across multiple hanger designs and for openly listing the limitations: grasping is simplified as attachment, hanger motion is kinematic, and the sim does not model gripper-garment contact in detail.\n\nThe soft spots are real but not disqualifying. The sim success criterion is a ray-cast coverage threshold (0.95 for left, 0.9 for right) set manually and never validated against real ground truth. If that threshold is too loose, the learned policies could be rewarded for “coverage” rather than true insertion, and the reported S=85.1% could be optimistic. But the real-world success is assessed independently, and the fact that the policy transfers to 8 unseen garments suggests the sim feedback is not drastically wrong. Still, the real evaluation is only 16 trials per condition with no error bars and no explicit protocol for judging success, so the 75% headline has a wide confidence interval. I would have liked to see at least a statement of how real success was judged, and ideally more trials or per-garment breakdowns. The hanger generalization results honestly show that performance drops with hangers lacking a crossbar or with high friction, which is useful but also limits the “generalizable” claim.\n\nWho is this for? Researchers working on deformable-object manipulation, sim2real for cloth, and household robotic tasks. The method is not foundational—the primitives are task-specific and the architecture follows prior work like ScissorBot and Bag All You Need—but the task decomposition and data-collection loop are sensible and reproducible enough to build on. The absence of released code or assets is a drawback, though not a fatal one.\n\nMy recommendation: this deserves a serious referee. It is a complete system with a new task, promising real-world evidence, and honest limitations. I would push the authors to report the real-world success-check protocol, add statistical confidence or more trials, and provide a sanity check that the sim coverage criterion matches human judgment on a handful of sim and real examples. If those are addressed, I would support conditional acceptance.","headline":"A solid sim-to-real pipeline for a genuinely underexplored task—hanger insertion into unseen garments—with a promising 75% real-world result, but the sim success metric and small real trial count keep the headline number provisional.","tokens_in":11759,"tokens_out":1734,"would_cite":true,"duration_ms":18163,"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":"This paper reports that a dual-arm robot can insert a hanger into unseen garments by learning four 2D keypoints from depth images and masks, trained entirely in simulation, achieving 75% real-world success without any real-data fine-tuning.","keywords":["hanger insertion","deformable object manipulation","sim-to-real transfer","dual-arm robot","action primitives","depth image policy","garment keypoints","cloth simulation"],"falsifier":"Test the trained policy, without any retraining, on a thin, slippery garment and a visibly flexible hanger—conditions that stress the simulator's simplified grasp model and rigid-hanger assumption. If insertion success falls far below the 75% reported on the original eight garments, the sim-to-real generalization claim is falsified.","tokens_in":10782,"feed_emoji":"🧥","tokens_out":9430,"duration_ms":77866,"temperature":0.7,"pith_summary":"Most robotic clothing research stops at flattening or folding; actually pushing a hanger through a garment's neckline is usually done by hand. This paper claims that this step can be learned: train a dual-arm robot in a cloth simulator on synthetic garments, treating the whole long sequence as two short motions — press-and-lift to insert the left end of the hanger, then drag-and-rotate to insert the right end — and let each motion be controlled by just two 2D points chosen from a depth image plus object masks. After training on 120 synthetic garments with no real-world demonstrations or fine-tuning, the policy succeeds in 75% of trials on eight unseen real garments, with an 85.1% success rate in simulation. The practical point is that a contact-rich, two-armed deformable-object skill, which prior robotic-hanging work avoided, may be tractable through low-dimensional action parameterization plus simulator-generated data.","feed_headline":"Trained in simulation, robot hangs unseen clothes 75% of the time","feed_subtitle":"No real-world demos or fine-tuning: depth-image keypoint policies transfer to eight unseen real garments.","key_machinery":"The load-bearing mechanism is the action-primitive parameterization: each subtask is a single short trajectory defined by two 2D keypoints — press-and-lift uses (press point, lift point) and drag-and-rotate uses (drag point, rotate point) — rather than a long sequence of joint angles. Four image-to-image networks are trained to output value maps whose peaks give these keypoints; the value at each pixel estimates the success probability of executing the primitive with that pixel as the target, so inference is an argmax over a 2D image. Conditioning the second network of each primitive on the first point (via a one-hot map) lets the two networks be trained cooperatively with binary cross-entropy losses. The second pillar is the simulator and data loop: an iterative scheme in which a heuristic policy seeds a dataset, the learned networks collect more trajectories, and the networks are fine-tuned on the combined set, all on synthetic garments and hangers rendered as depth and mask images.","core_discovery":"The paper's claim is that hanger insertion, despite its long horizon and high-dimensional cloth state, is solved by reducing it to four 2D keypoint predictions. The robot detects three keypoints on the hanger and three masks (hanger, garment, neckline). The first primitive, press-and-lift, takes a press point and a lift point: the hanger's left tip presses the neckline while the other gripper lifts the fabric, then the tip slides in. The second primitive, drag-and-rotate, takes a drag point and a rotate point: the robot re-grasps the hanger to avoid self-collision, drags the garment sideways, rotates the hanger so its right end aligns with the neckline, and pulls the garment back. Four networks output pixel-wise success maps; the highest-valued pixel in each map is the chosen keypoint, and the z-coordinate is held fixed to avoid depth-sensor noise. The networks are trained in a custom finite-element cloth simulator through an iterative loop of heuristic data followed by online self-collected data, using depth and masks as input to sidestep color and texture variation. The reported result is 85.1% overall success in simulation and 75% on eight unseen garments in the real world with no real-data fine-tuning.","pith_inferences":["The same two-point primitive scheme could be reused for other bimanual insertion-through-a-loop tasks, such as threading a belt through loops, inserting a shoelace, or hanging bags by their handles, where the essential problem is guiding one rigid endpoint through a deformable opening.","Because the policy reads depth and masks, it may be possible to couple it with a strong open-vocabulary segmentation model and a depth-refinement step in cluttered scenes; whether that preserves the 75% success rate is a testable extension of the paper.","The paper's success criterion in simulation is geometric ray coverage rather than physical contact; a real-world closed-loop verifier that confirms actual insertion could enable recovery from failures and improve the right-end insertion rate beyond 80%.","The stated limitation that gripper-garment contact is simplified suggests the biggest headroom for improving the method is a more realistic grasp model in simulation; if such a model raises the simulated success rate, the real-world rate would likely follow."],"forward_implications":["A long-horizon deformable manipulation task can be handled by four 2D keypoint predictions rather than dense joint-angle trajectories, which is why the policy transfers across garment types.","Because input is depth plus masks rather than color, the policy is insensitive to garment color and pattern; new garment geometries generalize if their depth and mask structure is similar.","The iterative online data collection supplies failure samples that improve the final success rate over heuristic-only data (85.1% vs 81.8% in simulation), and the neckline mask adds a further gain.","The success of the method is sensitive to hanger properties: in real-world tests, a hanger with a crossbar helps separate cloth layers, while a high-friction hanger wrinkles the garment and lowers success from 75% to 56.2%.","With human initialization the real-world success rises to 80%, indicating the policy is somewhat robust to how the garment is laid out."],"supporting_citations":[{"why":"supplies the garment-flattening initialization that places the garment flat on the table, the starting state of this task.","marker":"[2]"},{"why":"introduces action primitives as short parameterized end-effector motions, the control scheme the policy is built on.","marker":"[7]"},{"why":"provides the image-to-image network architecture used for the four keypoint value-map predictors.","marker":"[10]"},{"why":"supplies the GPU-parallel computation language underlying the custom cloth simulator used to generate training data.","marker":"[11]"},{"why":"performs the open-vocabulary segmentation of hanger, garment, and neckline masks that the policies take as input.","marker":"[24]"},{"why":"provides the large-step implicit integration method on which the simulator's cloth dynamics are based.","marker":"[29]"},{"why":"refines real-world depth images before policy execution to reduce the sim-to-real perception gap.","marker":"[31]"}],"fun_headline_variants":["Sim-trained robot hangs 8 unseen garments at 75%","Hanger insertion solved via keypoints, no real fine-tuning","From sim to reality: robot hangs clothes with 75% accuracy","RoboHanger: 75% real-world success on unseen garments","Depth-only keypoints let robot hang novel clothes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result stands on the assumption that cloth behavior in the custom simulator—where grabbing a garment is simplified to attaching it to the gripper and the hanger moves as a rigid body—is close enough to real physics that actions learned there work on a real robot.","fun_headline_variants_meta":{"raw":{"variants":["Sim-trained robot hangs 8 unseen garments at 75%","Hanger insertion solved via keypoints, no real fine-tuning","From sim to reality: robot hangs clothes with 75% accuracy","RoboHanger: 75% real-world success on unseen garments","Depth-only keypoints let robot hang novel clothes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1301,"prompt_tokens":1006,"completion_tokens":295,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":209}},"tokens_in":622,"tokens_out":295,"duration_ms":3497,"temperature":1.0,"reasoning_tokens":209,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:41:58.111712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the trained policy, without any retraining, on a thin, slippery garment and a visibly flexible hanger—conditions that stress the simulator's simplified grasp model and rigid-hanger assumption. If insertion success falls far below the 75% reported on the original eight garments, the sim-to-real generalization claim is falsified.","supporting_citations":[{"cited_title":"Cloth funnels: Canonicalized-alignment for multi-purpose garment manipulation,","cited_arxiv_id":null,"evidence_quote":"supplies the garment-flattening initialization that places the garment flat on the table, the starting state of this task."},{"cited_title":"Bag all you need: Learning a generalizable bagging strategy for heterogeneous objects,","cited_arxiv_id":null,"evidence_quote":"introduces action primitives as short parameterized end-effector motions, the control scheme the policy is built on."},{"cited_title":"Large steps in cloth simulation,","cited_arxiv_id":null,"evidence_quote":"provides the large-step implicit integration method on which the simulator's cloth dynamics are based."},{"cited_title":"D3RoMa: Disparity Diffusion-based Depth Sensing for Material-Agnostic Robotic Manipulation","cited_arxiv_id":"2409.14365","evidence_quote":"refines real-world depth images before policy execution to reduce the sim-to-real perception gap."}],"review_version":1}