{"id":"1b3fd7c8-9367-448a-8ad5-de3c0602901a","arxiv_id":"2501.15068","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data-driven framework decomposes manipulation tasks into reusable atomic skills, fine-tunes a VLA model per skill, and reports reduced data needs with comparable or better real-robot success rates.","lead":"This paper builds a library of reusable atomic skills for robots by having a vision-language planner break tasks into subtasks, then fine-tuning a vision-language-action model on small per-skill datasets. The authors report that their method matches or beats end-to-end training with less data and can recombine existing skills for new tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"New-task data-efficiency claim is untested: Table 2 only recombines already-trained skills; no experiment adds a genuinely missing skill, and abstraction accuracy is never measured.","rationale":"The paper's strongest claim is that skill decomposition substantially reduces data costs and enables efficient adaptation to new tasks. The adaptation claim depends on the ability of the VLP and semantic-abstraction modules to map subtasks of a new task onto an existing skill library and to create genuinely new skills when needed, with only that skill's data collected. The current experiments do not test this: Table 2 reuses the three block-moving skills that were already fine-tuned, so no missing skill is ever introduced. Therefore the core new-task mechanism is unvalidated. The reader's weakest_assumption identifies exactly this reliance on untested decomposition/abstraction quality. I agree with that identification. The reader's CONDITIONAL verdict is appropriate: the paper has a plausible system and some promising real-robot results, but the central data-efficiency claim for new tasks would be substantially strengthened by a single experiment that actually adds a missing skill and measures composed full-task success. Since my concern reinforces the reader's concern rather than moving the verdict in a different direction, the verdict should remain unchanged.","tokens_in":9436,"tokens_out":6013,"duration_ms":58106,"concrete_test":"Run the §3.1 'give the guest a cup of water' scenario: build a library containing only 'lift up the bottle' and 'align and tilt the bottle towards the cup'; introduce the new task requiring the genuinely missing skill 'deliver the cup'; collect demos only for that missing skill, fine-tune it, and evaluate the composed full-task success over at least 20 trials. Compare against an end-to-end VLA baseline trained on the full new task with the same total number of demos. Also have an independent annotator verify that the VLP agent identifies 'deliver the cup' as missing and that the semantic abstraction module produces a skill definition that is reusable across at least two distinct task variants.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a new task can be executed by composing existing atomic skills and, when a skill is missing, collecting data only for that skill (§3.1, §3.4). Table 2 is the only evidence for this adaptation claim, but every 'new task' is a permutation of the same three block-moving skills that were already trained; the system never encounters a genuinely missing skill. Consequently, the VLP and semantic-abstraction modules are never actually exercised in the scenario that matters most: they never have to recognize a missing skill, produce a new reusable skill definition, and confirm that the updated library composes correctly. The paper also reports no measure of decomposition or abstraction accuracy, so it is unknown whether the subtask-to-skill mapping yields consistent, task-independent skills. If the abstraction produces task-specific skill definitions, the library fragments and the claimed data-efficiency benefit for new tasks collapses. This is a correctness risk in the core mechanism, not a disagreement with consensus, and it is not resolved by the current experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-wheeled framework for building an atomic skill library for embodied manipulation. A VLP agent decomposes a task into subtasks; a semantic abstraction module maps these subtasks to reusable atomic skill definitions; and a VLA model is fine-tuned with small per-skill datasets to construct the library. The authors claim that this decomposition reduces data costs relative to end-to-end training, improves generalization to out-of-distribution object positions, and enables adaptation to new tasks by composing existing skills. Experiments are conducted on a real ALOHA dual-arm robot with two VLA backbones (Octo and RDT-1B) across four tasks, including a block-reordering task used for the new-task evaluation.","tokens_in":9608,"tokens_out":8162,"duration_ms":73910,"significance":"The problem is timely and the general direction is plausible: shifting from end-to-end task data to reusable skill-level data could indeed mitigate the data explosion faced by VLA policies. The paper reports real-robot experiments with two different VLA backbones, which is a strength, and the proposed dynamic skill-library concept goes beyond methods that rely on a fixed predefined skill set. If the central claims were supported by adequate evidence, the work would be a useful contribution to data-efficient manipulation. The current evidence, however, is too thin and confounded to establish the data-efficiency and adaptation claims.","major_comments":[{"comment":"The evaluation uses 10 trials per condition with no confidence intervals or significance tests. For binary success/failure outcomes, 10 trials cannot resolve the 10-20 percentage point differences that the Q1/Q2 analysis relies on; for example, a change from 40 to 60 is only two trials. The text in §4.2 even says 'All the success rates of our method are no less than the end-to-end method,' but in the Pour Water mug-OOD condition Octo(Ours) grasp is 40 vs Octo(End-to-end) 60. Please report confidence intervals, use more trials or statistical tests, and define a composite task success rate in addition to the per-phase rates.","section":"§4.1 / Table 1"},{"comment":"The 'Ours-plus' comparison is confounded. It is defined as maintaining the same data volume as end-to-end while using a 'larger distribution of data points,' so it changes both the data distribution and the use of the skill decomposition simultaneously. The observed gains (e.g., RDT(Ours-plus) banana both-OOD 80/70 vs RDT(Ours) 60/30) cannot be attributed to the atomic-skill library. A control that trains an end-to-end model on the same position-diverse data with the same total demo count is needed before claiming that skill decomposition, rather than data diversity, drives the improvement.","section":"§4.2 (Q2)"},{"comment":"Table 2 is the only evidence for the new-task adaptation claim, but every condition is a reordering of the same three block-moving skills that were already trained (red, green, blue). The scenario in §3.1 and §3.4 where a genuinely missing skill is detected, newly abstracted, learned from additional data, and then composed is never tested. Please add an experiment with a novel skill absent from the library, report the additional data required, and verify that the VLP agent recognizes the missing skill and that the updated library composes correctly on the new task.","section":"§3.1 / §3.4 / Table 2"},{"comment":"The manuscript never measures the quality of the VLP decomposition or the semantic abstraction of subtasks into atomic skill definitions. Since the data-efficiency and transfer claims depend on skills being reusable across tasks, the absence of any evaluation of decomposition or abstraction accuracy is a load-bearing gap. Please report examples of generated subtasks and skill definitions, a quantitative check of decomposition and abstraction consistency (e.g., human agreement or plan-execution success), and a clear statement of how skill granularity is set in practice.","section":"§3.2 / §3.4"},{"comment":"The data-cost accounting is incomplete. Section 3.3 states that the RDT-1B model was fine-tuned on 6,000 open-source plus 2,000 proprietary trajectories before the few-shot skill experiments, but the experimental demo counts (e.g., 9+9 for Pour Water) exclude this cost, and it is unclear whether the end-to-end baselines include it. The paper should state whether this 8,000-trajectory fine-tuning is a shared prerequisite for both methods and, if so, include it in the total-data-cost comparison.","section":"§3.3"}],"minor_comments":[{"comment":"The phrase 'exponential improvement of data collection efficiency' is not supported by the reported data counts, which show constant-factor reductions; please temper or quantify this claim.","section":"§1"},{"comment":"The data count for the Move blocks task is ambiguous: '10 demos of moving red, green, and blue block respectively' could mean 10 per color (30 total) or 10 total; please clarify.","section":"§4.1"},{"comment":"Table 1 reports per-phase success rates (e.g., 'Pick up | Place' and 'Grasp | Pour') but the text frequently refers to 'task success rate'; please define a composite success metric for multi-phase tasks or state explicitly that per-phase rates are reported.","section":"§4.1 / Table 1"},{"comment":"The comparison against end-to-end baselines in Table 2 is informative only as a zero-shot transfer test, since the end-to-end model is trained exclusively on one order and is given no new-task data; please state this more precisely rather than concluding that end-to-end methods 'cannot handle new tasks at all.'","section":"Table 2"},{"comment":"Some implementation details of the semantic abstraction module (e.g., prompts, output schema, and how skill granularity is chosen) are missing; providing these details would improve reproducibility.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The topic is well suited to the journal and the framework is plausible, but the empirical foundation is currently too thin for acceptance. The most important missing piece is an experiment that actually exercises the 'missing skill' mechanism of the library-update claim; without it, the central novelty of dynamic library expansion is not demonstrated. I would not reject the paper, since the core direction is defensible and the missing experiments are feasible within the existing setup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe one thing worth knowing: this is a real system paper, not a theory paper. The dynamic, self-updating atomic skill library is a genuine departure from RoboMatrix's predefined meta-skills, and the authors actually built it: VLP decomposition with Prismatic/DINO-X/SAM-2/GPT-4, semantic abstraction to skill definitions, then few-shot fine-tuning of RDT-1B or Octo on real ALOHA hardware. They run four real-world tasks, two backbone VLAs, and report success rates for in-distribution and out-of-distribution positions. The \"Ours-plus\" comparison—same data budget, wider position coverage—is a nice demonstration that skill decomposition lets you spend the same data on more diverse placements, and the block-ordering results show that composition of already-trained skills transfers to new orderings. The citation pattern is honest; they credit RDT-1B, Octo, RoboMatrix, and the perception tools they build on.\n\nNow the soft spots, in proportion. The central adaptation claim—that a new task needing a genuinely missing skill requires data only for that skill—is never actually tested. Table 2 recombines the same three block-moving skills that were trained; no experiment adds a new skill, so the semantic abstraction module is never exercised in the update scenario that is the paper's whole reason to exist. The paper also never measures decomposition or abstraction accuracy, so we don't know whether the subtask-to-skill mapping yields stable, task-independent definitions. Second, the data-efficiency numbers are more modest than the prose: \"exponential improvement\" appears in the intro, but the experiments show roughly 25–35% fewer demos (e.g., 27 vs 18 for pour water), not exponential scaling. Third, 10 trials per condition with no confidence intervals or significance tests is thin for the strong \"no less than\" comparisons; several differences are within one or two successes. Fourth, no code or data release, which limits reproducibility given the prompt engineering and rule-based spatial relation logic.\n\nNone of this is fatal to the core idea. The dynamic library mechanism is plausible, and the direction of the results is consistent with the claims. But the paper currently overstates what is shown, especially the missing-skill case and the \"exponential\" phrasing.\n\nWho benefits: researchers working on skill-based manipulation or VLA fine-tuning will want to know this exists and will want to see the missing-skill experiment. It deserves a serious referee, with the strong request that the authors either add that experiment or soften the claim, add statistical grounding, and release artifacts. I'd take it to a reading group—it's a good case study in how to (and how not to) evaluate a skill library.","headline":"A genuinely new dynamic skill-library idea with real-robot evidence that is weaker than the claims: the missing-skill adaptation scenario is never tested and the \"exponential\" data-efficiency claim is unmeasured.","tokens_in":10203,"tokens_out":2872,"would_cite":true,"duration_ms":27859,"reading_group":"yes","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 decomposing manipulation tasks into reusable atomic skills, then fine-tuning a vision-language-action model per skill, cuts data costs while matching or exceeding end-to-end success rates.","keywords":["atomic skill library","embodied manipulation","vision-language-action models","task decomposition","data efficiency","few-shot fine-tuning","robotic manipulation"],"falsifier":"A concrete test: assemble a new task entirely from skills already present in the library, but with the objects placed outside every distribution seen during skill training, and measure success; if the success rate is no better than for an end-to-end model that never saw the task, then the claimed cross-task transfer is not real. The paper's block-ordering results, which show 80-100% success on unseen orders, are the positive instance of this test.","tokens_in":9216,"feed_emoji":"🤖","tokens_out":5197,"duration_ms":44527,"temperature":0.7,"pith_summary":"The paper tries to establish that decomposing end-to-end manipulation tasks into atomic skills—small reusable actions like \"lift the bottle\" or \"align and tilt\"—cuts the data needed to train a robot manipulator while keeping or improving success rates. The proposed three-wheeled method uses a vision-language-planning agent to split a task into subtasks, a semantic abstraction step to consolidate those subtasks into general skill definitions, and few-shot fine-tuning of a vision-language-action model to realize each skill in an expandable library. The authors report real-world experiments on a dual-arm robot showing that with less data the skill-based method matches end-to-end training, with equal data it outperforms it, and that entirely new task orderings can be executed by recombining existing skills. If the claim holds, the path to general robot manipulation shifts from collecting full task demonstrations toward maintaining a growing library of reusable skills.","feed_headline":"Atomic skills cut robot data needs by reusing learned moves","feed_subtitle":"Decompose tasks into reusable skills; new tasks add data only for missing skills.","key_machinery":"The load-bearing object is the atomic skill library, built by a three-wheeled loop: the VLP agent wheel (GPT-4 prompted with scene descriptions, object bounding boxes from DINO-X, and segmentation masks from SAM-2 to plan subtasks), the VLA wheel (any pretrained vision-language-action model, e.g., RDT-1B or Octo, fine-tuned with few demonstrations per skill), and the atomic skill wheel (a semantic abstraction module, again an LLM, that maps diverse subtasks to a compact set of skill definitions at a granularity set by VLA plasticity and adaptability). The loop makes the library self-updating: when a new task needs a missing skill, only that skill's data is collected and fine-tuned.","core_discovery":"The central discovery claimed is that a robot policy trained per atomic skill—rather than per end-to-end task—yields a dynamically expandable library whose coverage of new tasks grows with each added skill, so the marginal data cost of a new task is at most the data for the skills it lacks. The paper's Table 1 shows skill-based fine-tuning matching or beating end-to-end fine-tuning on Pour Water, Pick & Place Banana, and Pick & Place Pen, with up to 40 percentage-point gains in out-of-distribution success under equal data budgets, and Table 2 shows that rearranging block-ordering subtasks succeeds at 80-100% without retraining on the new order, where end-to-end models fail entirely. The authors take this as evidence that mapping tasks to atomic skills materially reduces data requirements and enables cross-task generalization.","pith_inferences":["The paper's data-efficiency claim implicitly assumes that semantic abstraction yields skills that transfer across tasks; a direct test would measure whether a skill fine-tuned for one scene or object works in a new scene with zero or very few demonstrations.","If VLA plasticity is low, the required skill granularity becomes very fine, and the library may fragment into near-task-specific entries, eroding the data savings; evaluating this trade-off would require systematic variation of VLA capacity.","A natural extension is to treat skill reuse as the evaluation metric itself, for example by reporting the percentage of new tasks whose skills are fully covered by an accumulated library, which would quantify the avoidance of data explosion directly."],"forward_implications":["A new task whose subtasks are already covered by the library can be executed with zero new data collection.","The data needed to add a task scales with the number of missing skills, not the complexity of the full task.","Fixed data budgets can be spent on more diverse object and scene positions per skill, improving out-of-distribution generalization.","The framework is agnostic to the choice of VLA backbone, so improvements in pretrained policies directly upgrade the library.","As the library expands, the set of addressable tasks grows through skill recombination."],"supporting_citations":[{"why":"Supplies RDT-1B, the pretrained VLA backbone used for the main skill fine-tuning experiments.","marker":"[Liu et al., 2024]"},{"why":"Supplies Octo, the second VLA backbone used to show the method transfers across models.","marker":"[Team et al., 2024]"},{"why":"Provides the Mobile ALOHA dual-arm teleoperation system used to collect skill trajectory data.","marker":"[Fu et al., 2024]"},{"why":"Provides the Prismatic vision-language model used inside the VLP agent for scene description.","marker":"[Karamcheti et al., 2024]"},{"why":"Previous skill-centric framework limited to a manually predefined skill set; the paper contrasts its dynamic library against it.","marker":"[Mao et al., 2024]"},{"why":"Supplies DINO-X object detection used to compute spatial relations for task decomposition.","marker":"[Ren et al., 2024]"}],"fun_headline_variants":["Atomic skills let robots reuse moves, cutting data per task","Robot builds skill library on the fly, needs less data for new tasks","New robot method: learn atomic skills, expand library to handle more tasks","Skill-based robot training: add only missing skills for new tasks","Reusable robot skills cut data needs across tasks, study shows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the VLP agent and the semantic abstraction module can turn each subtask into atomic skill definitions that remain valid and reusable when the task changes, but the paper never measures decomposition or abstraction accuracy, so if skills are inconsistent or task-specific, transfer fails and the data-efficiency claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Atomic skills let robots reuse moves, cutting data per task","Robot builds skill library on the fly, needs less data for new tasks","New robot method: learn atomic skills, expand library to handle more tasks","Skill-based robot training: add only missing skills for new tasks","Reusable robot skills cut data needs across tasks, study shows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3170,"prompt_tokens":923,"completion_tokens":2247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":539,"tokens_out":2247,"duration_ms":15821,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:38:54.411161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: assemble a new task entirely from skills already present in the library, but with the objects placed outside every distribution seen during skill training, and measure success; if the success rate is no better than for an end-to-end model that never saw the task, then the claimed cross-task transfer is not real. The paper's block-ordering results, which show 80-100% success on unseen orders, are the positive instance of this test.","supporting_citations":[],"review_version":1}