{"id":"c8c7f034-9b6c-441d-b774-a7b532563c7f","arxiv_id":"2507.20832","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A neurosymbolic robot defines a mug's hanging part from its own support theory, collects training images, retrains its perception, and uses the part for planning.","lead":"This paper describes a robot architecture that combines neural object recognition with symbolic rules about support and contact, allowing the robot to define and learn a new object part concept, the part of a mug used when hanging from a hook. A generalist reader may care because it points to a path toward robots that teach themselves to see new functional parts instead of relying on a fixed, pre-programmed ontology.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No experimental section tests the claimed loop; the central claim that a detector trained on auto-labeled support contacts generalizes to new scenes and makes hanging-planning 'virtually guaranteed' is unverified.","rationale":"The paper is primarily a system description with a formal theory but no evaluation section. The central contribution is an empirical claim: an agent autonomously acquires a reusable functional-part concept and exploits it in planning. Since no success rates, detector accuracy figures, planning statistics, or comparisons are reported, the claim cannot be assessed. My concern aligns with the reader's weakest assumption about perception accuracy but is more specific: the auto-labeling loop uses the same symbolic Support predicates for generating training labels and for evaluating the resulting behavior, so the system can appear successful while merely rediscovering its own perceptual biases. The concrete test would settle this by measuring detector generalization on non-support frames and by comparing planning success against a whole-mug baseline. This does not alter the reader's verdict; rejection remains appropriate until an independent evaluation is provided.","tokens_in":10396,"tokens_out":4350,"duration_ms":51714,"concrete_test":"Run the full pipeline in the same simulator on a held-out set of mug and hook shapes and poses, including near-miss non-support configurations (mug bottom touching hook, tilted mug, hook behind body), and report: (1) precision/recall of the retrained MugSuppByHook detector against manually verified part labels on frames where no Support description holds; (2) success rate (mug remains hanging after release) for constraint solving with the detected part versus with the whole mug over at least 100 trials. The 'virtually guaranteed' claim requires near-100% part-based success and a statistically significant improvement over the whole-mug baseline; if either fails, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract and Section 5) is that the agent discovers the functional part MugSuppByHook from observations, retrains YOLO to recognize it, and uses it to plan stable support. The load-bearing premise is that the contact mask selected through Eq. 16 and formalized in Eq. 17 labels a repeatable force-bearing part, and that the retrained detector recognizes this part in novel, non-support contexts. That premise is untested at exactly the point where it carries the argument. The auto-labeled data are produced by the same perception predicates (contact, below, not moving down) and the same Support theory that would be used to judge success, so there is no independent ground truth; any bias in YOLO segmentation, optical flow, or contact masks is baked into the new class. Section 5's assertion that using the detected part makes support 'virtually guaranteed' additionally assumes the learned segmentation corresponds to the geometry that actually bears force during hanging. The paper reports no detector precision/recall, no planning success rates, and no baseline comparison. Without such measurements, the demonstrated loop is a described possibility, not a supported empirical result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neurosymbolic architecture for reactive robotics in which a first-order-logic theory of the Support image schema (Section 4) drives perception queries, diagnoses support situations from contact, relative-below, and non-downward-motion predicates, and defines a new functional part class MugSuppByHook (Eq. 17). The agent is claimed to use contact masks from support situations as automatic training labels, retrain a YOLOv8 detector to recognize this functional part in novel, non-support contexts, and then use the detected part to constrain a geometric constraint solver so that planning a mug-on-hook support becomes feasible. The abstract states that the approach is demonstrated in a simulated world.","tokens_in":10526,"tokens_out":9230,"duration_ms":102586,"significance":"If the claimed learning loop worked as described, the paper would make a useful contribution to neurosymbolic robotics and concept discovery: the idea of bootstrapping a perceptual class from an image-schematic theory and using it to focus planning is genuinely interesting, and the FOL theory in Section 4 is a coherent naive formalization of Support. The authors also provide a public source-code link. However, the manuscript reports no experiments, no quantitative metrics, and no comparisons; the central claim is an unverified proposal rather than a demonstrated result. Moreover, the auto-labeling procedure is self-referential to a degree that makes the claimed generalization unmeasurable without an independent test set. As submitted, the significance of the contribution cannot be assessed.","major_comments":[{"comment":"The paper's central claim—that the agent teaches itself to recognize a functional part and uses it to make hanging 'virtually guaranteed'—is not backed by any experimental results. The abstract announces a demonstration in a simulated world, but the manuscript contains no simulator description, no success rates for the planning loop, no precision/recall or accuracy for the retrained YOLO detector, and no comparison against a baseline. Without such measurements, the claimed 'demonstration' is a proposed mechanism, not a supported empirical result.","section":"Abstract; Section 5"},{"comment":"The learning loop is self-referential: MugSuppByHook is defined in Eq. (17) using DSupp and the perception predicates contact, below, and not moving down; the training labels are contact masks attached to situations selected by the same diagnostic rules; and success is judged by establishing a DSupp situation. There is no independent ground truth for what a handle is. Consequently, any systematic bias in YOLO segmentation, optical flow, or contact-region extraction is inherited by the new class, and the generalization claim ('recognize a MugSuppByHook even outside of a supported by Hook situation') requires a test set with manually annotated functional parts, which is not provided.","section":"Eq. (17), Section 5"},{"comment":"The diagnostic rule for DSupp is too weak: contact plus below plus not moving downward is also satisfied by an object resting on a table or shelf, so the auto-labeled data can mix hanging support with other support configurations. In addition, the consequent of Eq. (16) writes DSupp(e), applying the situation-description predicate to an object, whereas Eq. (15) applies DSupp(s) to a situation; as written the axiom is type-incoherent and cannot be used by the reasoner without correction.","section":"Eq. (16)"},{"comment":"The statement that solving the constraint for the detected functional part makes it 'virtually guaranteed that if a pose satisfying the constraint is found, then the mug is in fact supported by the hook' is not justified. Contact plus 'mug above' plus non-overlap is not a sufficient condition for static stability (friction, center of mass, and force balance are not modeled), and the paper gives no probabilistic analysis or simulation data to support the guarantee. The claim should be removed or replaced by a measured success rate.","section":"Section 5, 'virtually guaranteed'"}],"minor_comments":[{"comment":"Axiom 7 is missing the conjunction before ¬exrt(o, f2); the displayed formula is not syntactically well-formed.","section":"Eq. (7)"},{"comment":"There are several typos, e.g., 'Endorser of the ... claim' should be 'Endorsers' and 'Aspect such as' should be 'Aspects such as'; the text would benefit from copyediting.","section":"Section 1"},{"comment":"The paper states that some domain and range axioms are 'filled in by the reader'; for a formal theory this should be made explicit in an appendix or supplementary material, since the axioms as given are not a complete theory.","section":"Section 4"},{"comment":"The symbol for the functional part is written both as 'MugSuppByHook' and 'MugSuppbyHook'; please use one consistent name.","section":"Figure 5, Eq. (17)"},{"comment":"The contact mask is described informally as 'points near where this contact occurs'; the computation of 'near' and the mask extraction should be specified, preferably with parameters, to make the pipeline reproducible.","section":"Section 3.2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an extended abstract or position statement rather than a full archival paper. The absence of any evaluation is the main obstacle; I would consider a resubmission that includes a proper simulated-robot study with independently labeled test data and quantitative planning and detection results. There is no indication of misconduct; the issue is that the central empirical claim is unsupported as submitted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core idea is genuinely interesting: a robot that defines a functional part (MugSuppByHook) from its own symbolic theory of support, uses contact masks as training labels, retrains YOLO, and then uses the new class for constraint solving. That loop is not in the cited ISRL or image-schema papers. Second, the paper does not actually show it working. The abstract says 'we demonstrate in a simulated world,' but there is no experimental section, no success rates, no detector accuracy, no planning comparison. The demonstration is a description of what happens, not a measurement.\n\nWhat the paper does well: the FOL axioms in Section 4 are coherent as a naive theory of support, and Eq. 17 is a clean definition of a functional part. The architecture is described in enough detail to reproduce (there is a GitHub link for code and knowledge modeling). The discussion of Dreyfus's ontology trap is a nice framing for why an agent should be able to teach itself new perceptual categories. And the authors are honest about limitations: the final section notes that functional parts often become invisible during use and that annotations need to be done on frames where the function is not yet performed.\n\nThe soft spots are load-bearing. The central claim, that the trained network can recognize MugSuppByHook outside of supported situations and that this makes hanging 'virtually guaranteed,' is unverified. The training labels come from contact masks and the same Support predicates that judge success, so there is no independent ground truth; any bias in YOLO, optical flow, or contact detection is baked into the new class. The paper reports no precision/recall, no planning success, and no baseline. This is not a minor omission; it is the missing evaluation of the entire contribution.\n\nWho is this for: researchers working on neurosymbolic robotics, concept discovery, or affordances might find the idea useful. It is a solid idea paper, but it is not yet a solid system paper. As is, I would not accept it with results missing, but I would send it to review because the idea is worth referee attention. If the authors add even a simple simulation with success rates and generalization tests (e.g., new mug shapes, different hooks), it could become a good paper.\n\nRecommendation: send to peer review with major revision: require a proper experimental section.","headline":"Good idea, missing demonstration: the paper defines a functional-part discovery loop but never measures it.","tokens_in":11147,"tokens_out":2667,"would_cite":true,"duration_ms":29691,"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 robot learns to recognize handles by watching mugs hang from hooks","keywords":["neurosymbolic architecture","image schemas","functional parts","support relations","reactive robotics","concept discovery","perception grounding","ontology expansion"],"falsifier":"In the simulated world, take a mug in a support situation, extract the contact mask defined by axiom 16, retrain the detector on those masks, and then run the detector on scenes where the mug is not hanging to check that the detected MugSuppByHook region still coincides with the handle. If the region jumps between retraining runs or mismatches the handle, the claim that the network recognizes a functional part outside support situations fails.","tokens_in":10138,"feed_emoji":"🪝","tokens_out":4181,"duration_ms":50737,"temperature":0.7,"pith_summary":"This paper builds a robot agent that starts with no concept of a handle and, by observing mugs supported by hooks, invents the concept of a functional part: the part of an object that does the supporting work. The agent combines a neural vision module with an ontological theory of Support, so that perception queries, belief updates, and planning all flow through the same symbolic description. The central claim is that once the symbolic theory identifies the contact region in an observed support situation, the agent can collect those regions as training data, retrain its detector to recognize the part even outside support scenes, and use that part to plan hanging and unhanging actions. The significance is a concrete route out of the fixed-ontology trap for reactive robots: new perceptual categories arise from sensorimotor observation and become usable in planning.","feed_headline":"Robot learns to see handles by watching mugs hang on hooks","feed_subtitle":"A neurosymbolic loop turns observations of support into a new trained perception class, starting with no handle concept at all.","key_machinery":"The load-bearing mechanism is the image-schematic theory of Support coupled to a query-driven perception loop. Support is formalized using the Description-and-Situation pattern, with necessary roles supportee and supporter; axiom 16 turns observable symptoms, namely contact, supporter below, and supportee not moving down, into belief in a Support situation, while axiom 15 turns a believed Support description into perception queries for contact and movement relative to the floor. Contact masks returned by perception become the grounding data: axiom 17 defines MugSuppByHook as the part of a Mug that participates as supportee in a Support situation with a Hook, and those masks are stored as training examples. The same symbolic theory supplies constraints for a pose solver, so the discovered part simultaneously names a new perception class and focuses the planning search.","core_discovery":"The central claim is that an agent with no concept of a handle can acquire one by observing supported objects hanging from a hook, and can then use that concept in planning. The symbolic theory of Support states that a support situation exists when a supportee is in contact with a supporter, the supporter is below, and the supportee does not move down (Eq. 16). Applying that theory to perception outputs yields contact masks, and the agent defines a new class, MugSuppByHook, for the part of a mug that participates as supportee in such a situation (Eq. 17). The paper argues that the retrained neural network can recognize MugSuppByHook even outside a supported-by-hook scene, and that solving pose constraints with this part rather than the whole mug makes a satisfying pose virtually guarantee that the mug is actually supported by the hook.","pith_inferences":["A testable extension the paper leaves implicit is applying the same bootstrapping to other image schemas, such as Containment or Blockage, and measuring how much symbolic theory is needed to label data for each new functional part.","If the approach is ported to physical robots, the unmeasured quantity that will decide success is the stability of automatically selected contact masks; noisy real-world masks could make the new class drift between retraining runs.","The paper's 'virtually guaranteed' planning claim could be quantified as an unstable-pose rate: compare the fraction of constraint-solver solutions that fail to support when using the detected part versus the whole mug.","The architecture suggests a division of labor in which symbolic theories are deliberately small and only need to be accurate enough to label training data, not to model the world completely."],"forward_implications":["Because the agent creates its own labels from contact masks, it can expand its perception without human-annotated data for the new part.","The new class keeps its functional meaning: the network recognizes the part that can be a supportee in a hook support, not just a visual pattern.","Planning for support and support-destruction becomes tractable: searching over the handle part instead of the whole mug removes unstable poses that touch the hook without hanging.","The same mechanism can in principle be replayed for other image schemas, so the agent's ontology is open-ended rather than fixed at design time."],"supporting_citations":[{"why":"Supplies the Description-and-Situation pattern used to reify image schemas as situations with roles.","marker":"[3,4]"},{"why":"Defines how image schemas compose into complex events, motivating the frame-based situation modeling.","marker":"[10]"},{"why":"Introduces the image-schematic reasoning layer on which the agent's ontology module is based.","marker":"[26]"},{"why":"Provides the Semantic Reasoning Framework terminology and architectural structure the system inherits.","marker":"[32]"},{"why":"Supplies the YOLOv8 object detection networks used for pixel-level segmentation in the perception module.","marker":"[35]"}],"fun_headline_variants":["Robot discovers 'handle' concept by watching mugs on hooks","No handle concept? Robot learns it from hanging mugs","AI acquires handle perception via observed support relations","Robot gains handle recognition from watching support events","Neurosymbolic robot learns handle concept from observation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole loop depends on the vision module's contact masks and segmentations being accurate enough that the symbolic Support theory always picks out the same reliable physical part, and the paper reports no accuracy measurements for this.","fun_headline_variants_meta":{"raw":{"variants":["Robot discovers 'handle' concept by watching mugs on hooks","No handle concept? Robot learns it from hanging mugs","AI acquires handle perception via observed support relations","Robot gains handle recognition from watching support events","Neurosymbolic robot learns handle concept from observation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1290,"prompt_tokens":990,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":225}},"tokens_in":606,"tokens_out":300,"duration_ms":3817,"temperature":1.0,"reasoning_tokens":225,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:12:51.458215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the simulated world, take a mug in a support situation, extract the contact mask defined by axiom 16, retrain the detector on those masks, and then run the detector on scenes where the mug is not hanging to check that the detected MugSuppByHook region still coincides with the handle. If the region jumps between retraining runs or mismatches the handle, the claim that the network recognizes a functional part outside support situations fails.","supporting_citations":[{"cited_title":"Image schema combinations and complex events","cited_arxiv_id":null,"evidence_quote":"Defines how image schemas compose into complex events, motivating the frame-based situation modeling."},{"cited_title":"Thinking in front of the box: Towards intelligent robotic action selection for navigation in complex environments using image-schematic reasoning","cited_arxiv_id":null,"evidence_quote":"Introduces the image-schematic reasoning layer on which the agent's ontology module is based."},{"cited_title":"A survey of Semantic Reason- ing frameworks for robotic systems","cited_arxiv_id":null,"evidence_quote":"Provides the Semantic Reasoning Framework terminology and architectural structure the system inherits."},{"cited_title":"Ultralytics YOLO; 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv8 object detection networks used for pixel-level segmentation in the perception module."}],"review_version":1}