{"id":"f660b23b-cef4-4b40-8bfe-78da23c48ab5","arxiv_id":"2411.11714","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A three-level skill transfer framework combining LLM task planning, A* trajectory replanning, and tactile pose correction moved a drawer-opening-and-stacking skill to a cabinet scene with 8/10 real-robot success.","lead":"The paper builds a knowledge-graph skill library and uses it with an LLM, path planning, and tactile sensing to transfer a robot's drawer-opening and cup-stacking skill to a new cabinet-door scene. A smart generalist should read it because long-horizon, contact-rich robot manipulation is a core obstacle to useful home and factory robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-scenario demonstration without ablations leaves the 8/10 success unattributable to the proposed framework; central claim is under-supported.","rationale":"The central claim is an empirical validation: the framework transfers skills in a novel scene. The single most load-bearing assumption is therefore not any one component's correctness, but that the experimental outcome is attributable to the framework and generalizable. The paper's evidence fails this test on both counts: only one scenario, no ablations, and a baseline that cannot discriminate. The reader's weakest assumption (LLM planning) is plausible but secondary: the LLM demonstrably produced usable plans in 8/10 trials, and those trials succeeded; whether it will work in other scenes is a generalization question, not a falsification of the reported result. The most immediate threat is that the demonstration would survive removal of the framework's distinctive pieces. I therefore propose an ablation check. If the ablated variants still succeed, the paper's claimed contribution is not supported; if they fail, the result is meaningful, but the single-scenario evidence would still be too thin for the broad claim, leaving the conditional verdict unchanged. The tactile pose-estimation gap (the computation of q1 from contour points and Hough lines is never specified) strengthens the need for a control, because the physical-level contribution is otherwise unverifiable even in principle.","tokens_in":11271,"tokens_out":9407,"duration_ms":90540,"concrete_test":"Run the identical drawer-to-cabinet transfer on the same hardware with three ablations: (1) a fixed hand-written subtask sequence instead of GPT-4O planning; (2) straight-line interpolation instead of A*; (3) no tactile correction. Compare success rates over the same ten episodes. If any ablated variant achieves at least 7/10, the framework's components are not validated. Also record per-trial failure modes in the full framework to identify which level (task, motion, physical) caused the two failures.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim says the experiment validates the hierarchical skill transfer framework. That inference is load-bearing: if the 8/10 success is not actually caused by the framework's components, the central contribution collapses. The paper offers only one real-world transfer scenario (drawer-to-cabinet), no ablations, no scene variations, and no failure analysis. The sole baseline, direct SAC policy transfer from a different task, is a strawman that is expected to fail regardless of the proposed method. Consequently, the demonstration cannot distinguish the claimed contributions (LLM task-level transfer, A* trajectory adaptation, tactile pose correction) from manual engineering of the Neo4j graphs and prompt templates plus routine path planning. The Limitations section itself acknowledges the LLM's unreliable spatial reasoning, but the paper never reports how often the LLM produced incorrect sequences in the ten trials or whether such errors account for the two failures. Without a control that removes each component, the reader cannot attribute the positive results to the framework. This is the most load-bearing concern because it breaks the evidential link between experiment and conclusion, whereas the LLM reliability issue is a component-level risk that may or may not have affected the demonstration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical framework for transferring robot manipulation skills across similar scenes. A knowledge-graph skill library (task graph, scene graph, state graph) is used together with GPT-4O for task-level subtask-sequence transfer, an A*-based trajectory adapter for motion-level transfer, and a tactile-contour/pose-estimation module for physical-level adaptation. The experimental section reports a drawer-to-cabinet transfer scenario in which the proposed method achieves 8/10 success versus 0/10 for a directly transferred SAC policy, and a separate set of tactile contour-extraction RMSE comparisons against a fixed threshold.","tokens_in":11421,"tokens_out":1981,"duration_ms":21194,"significance":"If the central claim holds, the framework is a plausible example of how symbolic skill libraries, LLM reasoning, and tactile feedback can be combined for contact-rich, long-horizon manipulation without retraining. The paper has concrete strengths: it reports real-robot experiments, uses a physical tactile sensor, makes prompts and code available through a repository, and its tactile contour algorithm is evaluated on multiple geometric shapes. The main significance is therefore conditional: the claimed 8/10 success in an unseen scene would be a useful demonstration of compositional skill transfer, but the current evidence base is too narrow to establish that the framework, rather than manual engineering of graphs and prompts plus routine planning, is responsible for the result.","major_comments":[{"comment":"The central empirical claim rests on a single real-world transfer scenario (drawer to cabinet) with ten trials and no ablations, error bars, or confidence intervals. The manuscript states in Sec. IV-B that the robot 'successfully completed the task in the novel environment,' but with ten trials and no variance measure, the 8/10 success rate cannot be distinguished from chance variation or from the contribution of any single framework component. I request ablations that remove each layer (LLM task transfer, A* adaptation, tactile pose correction) and additional scene/task pairs, plus a failure analysis for the two unsuccessful trials.","section":"Sec. IV-B, Table I"},{"comment":"The only baseline, direct SAC policy transfer, is a strawman: training a policy on a drawer-opening task and applying it unchanged to a door-opening task is expected to fail because the action spaces and state distributions differ, and no adaptation mechanism is involved. This does not test whether the proposed hierarchical components add value over standard replanning or over simpler skill-retrieval baselines. The comparison should include, at minimum, a version of the proposed framework without the tactile layer and a version with only trajectory replanning, so that the 8/10 result can be attributed to the framework's specific mechanisms.","section":"Sec. IV-C, direct policy transfer baseline"},{"comment":"The paper admits in its Limitations that the LLM 'lacks a true understanding of the 3D physical world, which sometimes leads to inaccuracies in motion path generation.' Because the task-level transfer depends entirely on the LLM outputting a correct subtask sequence, the manuscript must report how often the LLM produced incorrect sequences across the ten trials, whether such errors caused the two failures, and what fallback or verification mechanism, if any, is used when the LLM output is wrong. Without this information, the robustness of the top layer of the framework is unquantified.","section":"Sec. V, Limitations; Sec. IV-B"},{"comment":"The tactile contour-extraction evaluation reports RMSE values for the proposed adaptive threshold and a fixed threshold, but it does not report the number of images per shape, standard deviations, or a statistical test. The claimed improvement for the 'line bundle' and 'general edge' cases is modest (e.g., 117.44 versus 214.16 for one condition), and without variance information it is unclear whether the differences are significant. Please report per-condition sample sizes, error bars, and a paired comparison, since the same sensor and shapes are used across methods.","section":"Sec. IV-D, Table II"},{"comment":"The definition of T_texture in Eq. (2) is ambiguous: the symbol G' is not defined, and the 'floor value of the average of non-zero gradient values' is not a standard quantity. Additionally, Algorithm 1 uses T_texture in the dual-threshold detection on line 12, but line 14 states that T_high and T_low are adjusted with T_texture; the relationship between these thresholds and their update rules is not specified. This makes the algorithm not fully reproducible from the paper alone, despite the claim of providing pseudocode.","section":"Sec. III-C, Eq. (2), Algorithm 1"}],"minor_comments":[{"comment":"The heading 'Implement Details' should be 'Implementation Details'.","section":"Sec. IV-A"},{"comment":"Reference [17] is described as 'Sun et al.' but the listed citation is Achituve et al.; the in-text author name should match the reference.","section":"Sec. II-B"},{"comment":"The overview figure is very dense and combines the framework, the three graph types, and an example in a single image; splitting the graph-construction and transfer-framework diagrams would improve readability.","section":"Fig. 2"},{"comment":"In the reward definition, the quantities 'success' and 'unsuccess' are not formally defined, and the reward values for the two cases are not stated as constants; please make the reward function explicit.","section":"Eq. (5)"},{"comment":"The notation |G'| is introduced without explanation; if it denotes the count of nonzero-gradient pixels, this should be stated explicitly.","section":"Sec. III-C, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central idea is interesting but the experimental validation is currently too thin to support the strong claims in Sec. IV-B. The lack of ablations, the weak baseline, and the unquantified LLM reliability are all fixable within the scope of a revision, so I do not recommend rejection. I would also encourage the editors to check that the linked repository contains the exact prompts and Neo4j graph definitions needed to reproduce the task-level transfer, since the paper relies on those details being public."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a legitimate integration of a knowledge-graph skill library, LLM task planning, A* motion replanning, and tactile pose correction, and it shows a real robot transferring a drawer-opening skill to cabinet-door opening. The 8/10 success is encouraging, but the evaluation is too thin to prove the framework components are responsible.\n\nWhat's new: the three-level decomposition itself, especially the 'state graph' with Require/Obtain as the bridge between task and scene graphs, is a clean way to feed scene information into an LLM planner. The tactile contour extraction algorithm with adaptive thresholds is evaluated separately (Table II) and shows consistent RMSE improvement over fixed thresholds. That part stands on its own as a useful component.\n\nSoft spots: the central experiment is one scenario, 10 trials, no failure analysis, and no ablations. The baseline—direct SAC policy transfer from a different task—is so weak that it mainly shows RL policies don't zero-shot transfer across articulation types. It doesn't tell you what the LLM, A*, or tactile correction each contributed. The paper itself states the LLM lacks real 3D understanding, but it never reports how often the LLM produced a bad plan or whether that explains the two failures. Code is linked but not verifiably shipped; I couldn't tell if prompts and trial logs are included.\n\nI don't see circularity: the knowledge graph is manually built and is an input, not a fitted parameter. The stress-test concern is fair but I'd phrase it as 'under-supported' rather than 'load-bearing flaw'—the framework is plausible, the evidence just doesn't isolate the mechanism.\n\nWho it's for: researchers doing system-building for manipulation skill transfer; it's a useful existence proof that these components can be stitched together. A serious referee should ask for more transfer scenarios, ablations, and release of prompts/logs. I'd send it to review rather than desk reject.","headline":"An honest integration paper whose real-robot demonstration is encouraging but whose evaluation is too thin to credit the framework's components.","tokens_in":12004,"tokens_out":2137,"would_cite":true,"duration_ms":20825,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a robot can transfer a long-horizon drawer-opening and cup-stacking skill to a new cabinet scene without retraining, with 8/10 success versus 0/10 for direct policy transfer.","keywords":["robot skill transfer","knowledge graph","skill library","task graph","scene graph","tactile perception","large language model","A* path planning"],"falsifier":"Run the same drawer-to-cabinet transfer on a scene where the door handle is at an orientation or height not represented in the scene graph, and check whether the LLM outputs a viable sequence; if the subtask sequence is wrong, execution fails before tactile correction. A more direct test is to repeat the ten-episode evaluation with the tactile pose-correction loop disabled: if success stays at 8/10, the physical-level module is not carrying the claimed load, and if it drops, the tactile representation is essential.","tokens_in":11021,"feed_emoji":"🤖","tokens_out":6864,"duration_ms":63606,"temperature":0.7,"pith_summary":"The paper claims that long-horizon, contact-rich manipulation skills can be transferred to similar but unseen scenes without retraining by organizing skill and scene knowledge as a knowledge graph and layering task, motion, and physical adaptation. The authors build a skill library with a task graph, a scene graph, and a state graph to connect high-level subtask plans with low-level object and joint information. An LLM transfers the subtask sequence, A* planning adapts trajectories, and tactile contour extraction corrects contact pose. In a drawer-to-cabinet transfer test, the framework succeeded 8/10 times while a directly transferred reinforcement-learning policy succeeded 0/10. The significance would lie in replacing scene-specific retraining with structured knowledge plus local tactile feedback.","feed_headline":"Drawer-trained robot opens cabinet door: 8/10 success","feed_subtitle":"Task graph, scene graph, and tactile feedback carry the transfer; a directly transferred RL policy scores 0/10.","key_machinery":"The load-bearing object is the knowledge-graph skill library, specifically the triad of task graph, scene graph, and state graph. The task graph decomposes a long-horizon task into subtasks and action primitives; the scene graph represents object parts, Cartesian poses, quaternion orientations, and joint attributes encoded in XML; the state graph connects the two through require and obtain relations, so high-level plans can query and update low-level scene state. This graph is what lets the LLM reason about transfers, what supplies start and end points and obstacles to A* trajectory planning, and what the tactile quaternion correction loop, $q_w = q_3 \\otimes q_2 \\otimes q_1$ and $q_e = q_t \\otimes q_w^{-1}$, plugs into.","core_discovery":"On the paper's own terms, the central discovery is that skill transfer across similar manipulation scenarios can be achieved by decomposing knowledge into semantic, geometric, and physical layers rather than by transferring a monolithic policy. A task graph holds the subtask sequence learned in the source scene; a scene graph holds object parts, poses, and joint constraints of the target scene; a state graph lets each subtask require scene state and obtain updated state after execution. At run time, an LLM prompted with the skill library in code form outputs the new subtask sequence, A* with a collision-loss term replans the trajectory, and a tactile pipeline extracts contours and computes a quaternion pose error to drive corrective motion. The paper's headline evidence is 8/10 successful task completions in the new scene versus 0/10 for a directly transferred SAC policy, reported in Table I.","pith_inferences":["Beyond the paper: the layered design implies a modular failure profile, where a wrong LLM subtask sequence fails before the motion or tactile layers run, so an automated verifier on the output sequence could catch most errors without retraining.","Beyond the paper: the tactile threshold is derived from the gradient distribution of the current image, which suggests it will degrade on very low-texture or saturated contact patches; an explicit test is to run the extraction on a blank or glossy surface and measure whether the contour set collapses.","Beyond the paper: the success-rate gap versus direct policy transfer is unsurprising given that the baseline receives no semantic knowledge, so a stronger comparison would be an LLM-based planner without the graph, isolating the contribution of the skill library itself.","Beyond the paper: if the graph encoding were made editable by non-experts, the same architecture could support rapid re-targeting to new household objects by adding nodes and joint tags rather than collecting demonstrations."],"forward_implications":["If the 8/10 success transfers to other similar scene pairs, the framework offers a recipe for retraining-free skill reuse: encode once in graph form, replan with heuristics, and close contact errors with tactile feedback.","The same library can express prismatic drawer motions and revolute door motions through XML joint attributes, so skill transfer should extend to articulated objects whose joint types and axes are known.","The tactile loop lets the same skill adapt to cups at different orientations, meaning precise stacking does not require a fixed grasp or a visual pose model.","Because the task layer is prompt-based, updating a skill library entry is a graph edit rather than a policy retrain, which would make incremental skill accumulation cheaper if the method scales."],"supporting_citations":[{"why":"Supplies the heuristic A* planner used for motion-level trajectory transfer and collision avoidance.","marker":"[12]"},{"why":"Frames the tactile representation on which the physical-level contour extraction and pose correction is built.","marker":"[13]"},{"why":"Grounds the knowledge-graph construction method behind the task, scene, and state graphs.","marker":"[27]"},{"why":"Supplies the SAC algorithm whose directly transferred policy is the 0/10 baseline in Table I.","marker":"[29]"}],"fun_headline_variants":["Semantic-geometric-physical skill transfer: 8/10","Graph-based skill transfer: 8/10 success, direct RL 0/10","LLM + tactile skill transfer: 8/10 in new scenes","Skill library + tactile: 8/10 transfer, 0/10 direct RL","From drawer to cabinet: 8/10 via skill transfer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline leans on the assumption that the LLM, prompted with a manually encoded graph and a fixed four-stage prompt, will produce a correct and executable subtask sequence for the new scene, because the paper itself notes the LLM lacks true 3D understanding and can generate inaccurate motion paths.","fun_headline_variants_meta":{"raw":{"variants":["Semantic-geometric-physical skill transfer: 8/10","Graph-based skill transfer: 8/10 success, direct RL 0/10","LLM + tactile skill transfer: 8/10 in new scenes","Skill library + tactile: 8/10 transfer, 0/10 direct RL","From drawer to cabinet: 8/10 via skill transfer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001388,"raw_usage":{"total_tokens":5639,"prompt_tokens":989,"completion_tokens":4650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":4551}},"tokens_in":605,"tokens_out":4650,"duration_ms":31074,"temperature":1.0,"reasoning_tokens":4551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:13:03.914438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same drawer-to-cabinet transfer on a scene where the door handle is at an orientation or height not represented in the scene graph, and check whether the LLM outputs a viable sequence; if the subtask sequence is wrong, execution fails before tactile correction. A more direct test is to repeat the ten-episode evaluation with the tactile pose-correction loop disabled: if success stays at 8/10, the physical-level module is not carrying the claimed load, and if it drops, the tactile representation is essential.","supporting_citations":[{"cited_title":"Improved a-star algorithm for robot path planning in static environment,","cited_arxiv_id":null,"evidence_quote":"Supplies the heuristic A* planner used for motion-level trajectory transfer and collision avoidance."},{"cited_title":"Robocse: Robot common sense embedding,","cited_arxiv_id":null,"evidence_quote":"Grounds the knowledge-graph construction method behind the task, scene, and state graphs."}],"review_version":1}