{"id":"5935d78d-1b79-4302-a157-2091696391b9","arxiv_id":"2507.18276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AdaRPG uses GPT-4o, GroundingDINO, and SAM to locate and segment the movable part, a part-affordance model to choose a grasp, and GPT-4o to write the control loop, outperforming prior methods on new articulated objects.","lead":"Robot manipulators need to open and close objects with moving parts, like doors, bottles, and pressure cookers, even when the internal mechanism is hidden. This paper combines AI models that find and describe the movable part with a small custom model that picks the best grasp point, and reports high success rates on objects the system has never seen.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Category overlap between training and test inflates the cross-category claim: five of seven AdaManip test categories (Bottle, Pen, Window, Door, Coffee Machine) appear in the PartNet-Mobility affordance training set, contradicting Sec. 4.1's 'entirely novel object categories.'","rationale":"The reader's weakest assumption correctly identifies the train/test category overlap as the central security risk, and I agree with that assessment. The manuscript's own Sec. 4.1 statement that the framework is tested on 'entirely novel object categories' is factually inconsistent with the category lists in Tables 1 and 4, so this is not an external concern but an internal overclaim. Weighing it fairly, the paper still has genuine support: the pressure cooker and lamp categories are absent from the PartNet-Mobility training list, and the method achieves high success there (1.00 and 0.70 in simulation, 10/10 and 8/10 in the real world), plus the w/o-affordance ablation shows the affordance model contributes meaningfully. These facts keep the result credible but conditional. The missing trial counts and confidence intervals are also noted but are secondary; the category overlap is the load-bearing issue because it directly conditions the central generalization claim. A leave-one-category-out retraining experiment would settle whether the improvement on overlapping categories is real part-level transfer or category memorization. Therefore the reader's CONDITIONAL verdict remains appropriate, with no verdict change needed.","tokens_in":12833,"tokens_out":4108,"duration_ms":42992,"concrete_test":"Retrain the affordance model on the same PartNet-Mobility part dataset after removing all parts from Bottle, Pen, Window, Door, and Coffee Machine, then rerun the Table 4 evaluation on those five categories. If success rates drop by more than roughly 10 percentage points from the reported 0.84, 0.73, 0.80, 0.84, and 0.78, the claimed cross-category generalization is substantially inflated by train/test category overlap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is cross-category generalization on novel articulated objects. The most load-bearing condition is the novelty separation between the part-affordance training data and the evaluation objects. Table 1 lists training categories Window, Bottle, CM, Door, Pen, Safe, Faucet, Pot, Microwave, Switch, and Toaster; Table 4 evaluates Bottle, Pen, PC, CM, Window, Door, and Lamp. Five of seven test categories are also training categories, directly contradicting Sec. 4.1's assertion that the framework 'is tested on entirely novel object categories.' Although only part point clouds are used in training, parts drawn from the same object categories (bottle caps, door handles, window handles, pen tips, coffee-machine buttons) share substantial geometric similarity with test parts, so the affordance model may exploit category-level part geometry rather than demonstrating true part-level generalization. The genuinely novel categories are PC and Lamp, and those results (sim 1.00 and 0.70; real 10/10 and 8/10) are strong evidence for some cross-category ability. However, they do not justify the broad seven-category generalization claim as stated. If the five overlapping categories were removed from training and success rates dropped materially, the headline result would be substantially inflated by category overlap rather than part-level generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaRPG, a framework for adaptive articulated-object manipulation that combines foundation-model-based part grounding (GPT-4o, GroundingDINO, SAM) with a part-level affordance model trained on a newly annotated PartNet-Mobility-derived dataset, and GPT-4o-generated Python control code that invokes primitive manipulation skills. The central claim is that the part-level representation enables cross-category generalization to 'entirely novel' articulated object categories. The paper reports simulation success rates across seven categories (Tab. 4) and real-world success rates across four object types (Tab. 5), consistently outperforming baselines such as SAGE, CoPa, and AdaManip, as well as two ablations.","tokens_in":13138,"tokens_out":3118,"duration_ms":35879,"significance":"If the claims are correct, the paper makes a useful contribution: it shows that a modular pipeline of frozen foundation models for part grounding, a lightweight part-affordance network, and LLM-generated adaptive control code can solve long-horizon manipulation tasks without training on the test objects. The part-level affordance dataset is also a potentially reusable resource. The strongest evidence is the performance on pressure cooker and lamp, which are not among the training categories; both simulation and real-world results on those categories are markedly better than the baselines. However, the breadth of the cross-category generalization claim is not supported by the current evaluation because five of the seven simulation categories and one of the four real-world categories overlap with the training categories. The paper also omits statistical detail (trial counts, seeds, variance) for the simulation experiments, making it difficult to assess the reliability of the reported margins.","major_comments":[{"comment":"The statement in Sec. 4.1 that the framework 'is tested on entirely novel object categories' is directly contradicted by the paper's own tables. Table 1 lists Window, Bottle, CM, Door, Pen, Safe, Faucet, Pot, Microwave, Switch, and Toaster as the source categories of the part-affordance training data, while Table 4 evaluates on Bottle, Pen, PC, CM, Window, Door, and Lamp. Five of the seven simulation test categories (Bottle, Pen, CM, Window, Door) are also training categories, and the real-world test set in Table 5 includes Microwave, another training category. Since the central claim is cross-category generalization, this overlap is load-bearing. The authors should either (a) provide object-level evidence that the test instances are disjoint from all training instances and that no part geometry from the same categories is shared, (b) re-run the evaluation with training categories fully disjoint from test categories, or (c) explicitly restrict the 'novel category' claim to PC and Lamp and present the overlapping categories as within-category or near-category generalization. As written, the reported gains on the overlapping categories may be inflated by category-level part similarity rather than true part-level generalization.","section":"Sec. 4.1, Tab. 1, Tab. 4"},{"comment":"The simulation results are reported as single point estimates with no number of trials, no seeds, and no measure of variance (standard deviation, confidence interval, or per-seed range). For a policy whose generated code contains stochastic elements (e.g., the np.random.rand() condition in Fig. 4) and whose perception pipeline itself may be stochastic, this makes it impossible to determine whether the reported margins over baselines and ablations are statistically meaningful. The paper should report the number of episodes per category, the number of random seeds, and the variance across seeds for success rate, affordance F1, and segmentation IoU.","section":"Tabs. 3 and 4"},{"comment":"The evaluation environment AdaManip [49] is authored by the same group, and the part-affordance training data are extracted from PartNet-Mobility, the same underlying dataset family used to construct many articulated-object benchmarks. The paper states that 'our affordance model is not trained on the AdaManip dataset,' but it does not establish that the AdaManip test objects are distinct object instances from the PartNet-Mobility objects used to extract training parts. If the test objects or their parts are drawn from the same source models, the affordance model could be memorizing part geometry rather than generalizing. The authors should specify the exact source of each AdaManip test object and provide a formal train/test split at the object-instance level, not merely at the category level.","section":"Sec. 3.1 and Sec. 4.1"},{"comment":"The real-world results, while encouraging, are reported as counts out of 10 with no trial protocol details (e.g., number of distinct physical objects per category, initial configurations, whether successes are all-or-nothing task completion or partial progress). This makes it hard to compare against the simulation numbers and to assess how much of the 'real-world performance surpasses simulation' observation is due to task difficulty differences rather than domain-gap effects. The authors should describe the real-world task definitions, the number of unique objects per category, and the success criteria in enough detail that the experiments could be reproduced.","section":"Sec. 4.4, Tab. 5"}],"minor_comments":[{"comment":"The segmentation IoU values in Table 2 are reported without any indication of the number of images or objects evaluated, nor any per-instance variance. Adding this information would strengthen the claim that grounding is robust.","section":"Sec. 4.1, Tab. 2"},{"comment":"The probability condition 'np.random.rand() < 0.1*step' in the generated code is a hand-tuned-looking schedule; the paper should clarify whether this schedule is produced autonomously by GPT-4o from the general prompt, or whether the prompt or post-processing enforces this specific form. This matters because the adaptive behavior is partly a property of this schedule.","section":"Sec. 3.4, Fig. 4"},{"comment":"The annotation procedure says 'high-affordance points are automatically computed based on the part's center and bounding box,' but the exact rule is not specified. A precise definition (e.g., radius, kernel, or threshold) would make the dataset construction reproducible.","section":"Sec. 3.1"},{"comment":"The related-work discussion of AdaManip states that it 'introduces five distinct adaptive mechanisms,' but the experiments here use seven categories; the relationship between the mechanisms and the chosen categories is not explained, and it would help the reader to know which mechanisms appear in which test category.","section":"Sec. 2.1"},{"comment":"The real-world section says the pipeline selects 'the point with the highest affordance score,' while the simulation section (Sec. 3.3) uses a threshold epsilon and averages over selected points. The discrepancy should be acknowledged or reconciled.","section":"Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The category-overlap issue is the main substantive problem, and it is fixable: the authors can rerun with fully disjoint categories, or reframe the claim and provide object-level disjointness evidence. The lack of variance reporting is also fixable. I do not see an unfixable flaw in the core idea; the pressure-cooker and lamp results suggest genuine cross-category ability. The use of an in-group evaluation environment (AdaManip) is a concern for impartial benchmarking, but not grounds for rejection if the authors provide object-level train/test separation and independent or clearly specified object sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline: the method is a reasonable modular recipe and the results are strong, but the central 'novel category' claim does not survive contact with the paper's own tables. Training categories (Tab 1) include Bottle, Pen, Window, Door, and Coffee Machine; evaluation categories (Tab 4) include the same five plus Pressure Cooker and Lamp. Sec 4.1 says the framework is tested on 'entirely novel object categories' — that is not accurate. The genuinely held-out categories are PC and Lamp, and there the results (1.00 and 0.70 sim, 10/10 and 8/10 real) do show some cross-category ability, so the method is not a wash. But the broad seven-category claim is inflated by category overlap.\n\nWhat is actually new: the composition of GPT-4o part description, GroundingDINO box, SAM mask, part point cloud, PointNet++ affordance, and GPT-4o-generated control code with an adaptive loop. The part-level affordance dataset is a new artifact, and the ablations (w/o prompt, w/o affordance) show each component matters. Real-world transfer from simulation-trained affordance to a Franka is credible.\n\nSoft spots, in order: (1) The category overlap is load-bearing and must be fixed — either retrain the affordance model without the overlapping categories, or re-frame the claim to per-part generalization from held-out parts rather than held-out categories. (2) No seeds, trial counts, or variance are reported for simulation; given the large deltas over baselines that is a minor annoyance, not a fatal flaw. (3) The rotation-to-pull probability schedule (0.1*step) is a hand-set heuristic in the generated code; no sensitivity analysis. (4) The evaluation environment AdaManip is from the same group; not a problem by itself, but it means the baseline comparisons are not independent. (5) No code or data release.\n\nWho this is for: robotics folks working on articulated object manipulation and foundation-model control. It is a practical recipe paper, not a theoretical breakthrough.\n\nRecommendation: deserves a serious referee. The evaluation claims need revision before acceptance; if the author can show results on genuinely held-out categories (or retrain without overlap), this is a solid venue paper. As it stands, conditional.","headline":"Category overlap undermines the headline generalization claim, but the method is a solid modular recipe with credible results on genuinely held-out categories.","tokens_in":13664,"tokens_out":2592,"would_cite":false,"duration_ms":23939,"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":"AdaRPG claims that segmenting an articulated object into functional parts and scoring each part's surface with an affordance network makes manipulation skills transfer to novel object categories.","keywords":["articulated object manipulation","affordance learning","foundation models","part grounding","cross-category generalization","adaptive manipulation","code-as-policies","point cloud"],"falsifier":"Run the same AdaRPG pipeline on articulated objects whose functional parts come from categories absent from Table 1 (for example, a folding ladder, an umbrella, or a sliding shower door) and compare success rates; alternatively, compute the geometric nearest-neighbor distance between each AdaManip test part and the PartNet-Mobility training parts and check whether success falls as that distance grows.","tokens_in":1804,"feed_emoji":"🤖","tokens_out":2402,"duration_ms":70305,"temperature":0.7,"pith_summary":"AdaRPG argues that the main obstacle to manipulating never-seen articulated objects is not the object as a whole but its individual functional parts. It builds a pipeline where GPT-4o describes and grounds the relevant part, GroundingDINO and SAM segment it, and a part-level affordance model trained only on detached part point clouds chooses where to grasp and act. Then GPT-4o writes Python control code that calls primitive skill functions such as grasp, pull, push, and rotate in an adaptive retry loop. The paper reports success rates of 0.70 to 1.00 across seven simulated categories and 8 to 10 out of 10 on four real-world objects, outperforming methods that train on all test categories or reason over whole objects. The central claim is that part geometry is similar enough across categories to make affordance transfer work, so a single reusable recipe can replace per-category policies.","feed_headline":"Robot opens unfamiliar objects by acting on parts, not whole objects","feed_subtitle":"A part-level affordance model plus foundation-model code generation reaches 8-10/10 on real-world articulated objects.","key_machinery":"The load-bearing object is the part-level affordance model and its dataset \\(\\mathcal{D}_{\\text{pa}} = \\{(O_i, p_i, r_i)\\}\\), where \\(O_i\\) is a part point cloud, \\(p_i\\) is a point, and \\(r_i\\) is a binary actionability label. The model is a PointNet++ encoder followed by an MLP and sigmoid, trained with binary cross-entropy loss \\(\\mathcal{L}_V = \\mathrm{BCELoss}_{\\mathcal{D}_{\\text{pa}}}(r_i, V(O_i, p_i))\\). At inference, the pipeline is a grounding chain: GPT-4o writes a detailed description of the actionable part, GroundingDINO converts that description to a bounding box, SAM segments the mask, the mask is back-projected onto depth to build a part point cloud, and the affordance model emits per-point scores that select a grasp pose. The six atomic skill functions operate in the end-effector frame under impedance control and return success flags, which the LLM-generated Python code uses in adaptive retry loops.","core_discovery":"The paper's central claim is that adaptive manipulation of novel articulated objects can be decomposed into three reusable pieces: grounding the object's functional part with frozen foundation models, scoring that part's surface with an affordance network trained only on detached part point clouds, and having an LLM compose primitive functions into an adaptive program. It reports that this decomposition lifts simulation success rates to 0.84, 0.73, 1.00, 0.80, 0.84, 0.78, and 0.70 across bottle, pen, pressure cooker, coffee machine, window, door, and lamp in the AdaManip environment, and 9/10, 10/10, 9/10, and 8/10 on a real pressure cooker, microwave, bottle, and lamp. The paper attributes the gains to the fact that parts share greater local geometric similarity than whole objects, so a part-level affordance model generalizes across categories that a whole-object model cannot.","pith_inferences":["Editorial inference: a cleaner test of the paper's premise would train the affordance model on part categories disjoint from the test categories in Table 4; the current overlap between Tables 1 and 4 leaves open whether the gains are part-level or category-level.","Editorial inference: the generated code's pull attempt becomes more likely as rotation count increases, which is a stochastic proxy rather than a direct observation of lock state, so replacing it with a learned state estimator could improve mechanisms with multiple hidden states.","Editorial inference: because the affordance dataset stores parts without full-object context, it can be reused for other downstream skills such as pressing, twisting, or sliding by re-annotating the affordance surface plane, giving the dataset a life beyond the six primitive functions.","Editorial inference: all perception and reasoning models are frozen, so AdaRPG's ceiling is set by pretrained segmentation and code-generation quality rather than by the amount of robot experience collected, which means the approach should improve automatically as those foundation models improve."],"forward_implications":["A single part-affordance model, trained once on detached part clouds, transfers across categories and environments without retraining, so new object types can be absorbed by updating only the grounding and code-generation prompts.","Because the ablation without GPT-4o part descriptions drops success to near zero, reliable open-vocabulary part grounding is a first-order requirement for the whole pipeline, not a cosmetic addition.","The affordance-guided contact selection adds roughly 15 percentage points of average success over using the part center, confirming that part-level actionability scoring is doing the work.","LLM-generated control loops that retry rotation and periodically test a pull can express hidden-state adaptation without task-specific scripts, allowing the same prompt to drive all seven simulated categories.","The method outperforms an imitation policy trained on all seven test categories while itself never seeing test objects, suggesting part-level priors can beat category-level demonstrations."],"supporting_citations":[{"why":"Provides the AdaManip simulation testbed with adaptive mechanisms and the imitation-learning baseline it must outperform.","marker":"[49]"},{"why":"Supplies the whole-object affordance formulation and PointNet++ training recipe that the part-level model adapts.","marker":"[38]"},{"why":"Serves as the whole-object affordance baseline for F1 comparison and informs the affordance learning loss.","marker":"[50]"},{"why":"Baseline that maps semantic parts to actionable parts with a VLM; compared in simulation and real-world tables.","marker":"[13]"},{"why":"Baseline that generates 6-DoF poses from part grounding; compared as a foundation-model part-based method.","marker":"[17]"},{"why":"GPT-4o system card; the model writes the part descriptions and the high-level Python control code.","marker":"[21]"},{"why":"GroundingDINO open-set detector converts the part description into a bounding box for segmentation.","marker":"[32]"},{"why":"SAM converts the bounding box into a precise part mask that is back-projected to a point cloud.","marker":"[23]"},{"why":"PartNet-Mobility supplies the source articulated objects whose parts form the affordance training dataset.","marker":"[37]"},{"why":"PointNet++ is the encoder backbone of the part affordance model.","marker":"[41]"}],"fun_headline_variants":["Part-level affordance empowers robots on unfamiliar articulated objects","Robots open novel articulated objects via part grounding and LLM reasoning","AdaRPG: foundation-model part reasoning for adaptive manipulation","Ground parts, reason in code: robots master unseen articulated objects"],"cache_read_input_tokens":15744,"weakest_assumption_plain":"The evaluation treats the test objects as novel even though several of the affordance-model training categories are the same as the test categories; if those test objects closely resemble training parts, the reported gains could come from category similarity rather than true part-level generalization.","fun_headline_variants_meta":{"raw":{"variants":["Part-level affordance empowers robots on unfamiliar articulated objects","Robots open novel articulated objects via part grounding and LLM reasoning","AdaRPG: foundation-model part reasoning for adaptive manipulation","Ground parts, reason in code: robots master unseen articulated objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000863,"raw_usage":{"total_tokens":3736,"prompt_tokens":932,"completion_tokens":2804,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2735}},"tokens_in":548,"tokens_out":2804,"duration_ms":21657,"temperature":1.0,"reasoning_tokens":2735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:15:45.825639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same AdaRPG pipeline on articulated objects whose functional parts come from categories absent from Table 1 (for example, a folding ladder, an umbrella, or a sliding shower door) and compare success rates; alternatively, compute the geometric nearest-neighbor distance between each AdaManip test part and the PartNet-Mobility training parts and check whether success falls as that distance grows.","supporting_citations":[{"cited_title":"Adamanip: Adaptive articulated object manipulation envi- ronments and policy learning","cited_arxiv_id":null,"evidence_quote":"Provides the AdaManip simulation testbed with adaptive mechanisms and the imitation-learning baseline it must outperform."},{"cited_title":"Guibas, Mustafa Mukadam, Abhi- nav Gupta, and Shubham Tulsiani","cited_arxiv_id":null,"evidence_quote":"Supplies the whole-object affordance formulation and PointNet++ training recipe that the part-level model adapts."},{"cited_title":"V AT-mart: Learning visual action trajectory proposals for manipulating 3d ARTiculated ob- jects","cited_arxiv_id":null,"evidence_quote":"Serves as the whole-object affordance baseline for F1 comparison and informs the affordance learning loss."},{"cited_title":"Sage: Bridging semantic and actionable parts for generalizable articulated-object manipu- lation under language instructions, 2023","cited_arxiv_id":null,"evidence_quote":"Baseline that maps semantic parts to actionable parts with a VLM; compared in simulation and real-world tables."},{"cited_title":"Copa: General robotic manipulation through spa- tial constraints of parts with foundation models","cited_arxiv_id":null,"evidence_quote":"Baseline that generates 6-DoF poses from part grounding; compared as a foundation-model part-based method."},{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"SAM converts the bounding box into a precise part mask that is back-projected to a point cloud."},{"cited_title":"Chang, Li Yi, Subarna Tripathi, Leonidas J","cited_arxiv_id":null,"evidence_quote":"PartNet-Mobility supplies the source articulated objects whose parts form the affordance training dataset."}],"review_version":2}