{"id":"640618ca-e868-4dc7-81d2-4d8da972440a","arxiv_id":"2508.21501","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A neuro-symbolic system learns symbolic task rules and neural control policies from as few as five demonstrations and generalizes to larger unseen task instances.","lead":"This paper builds a robot training method that learns both the plan and the control skills from just a few human demonstrations. It could make teaching robots long tasks simpler by auto-discovering the structural rules behind a task instead of hand-coding them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The symbolic generalization claim rests on an unverified identifiability assumption: Section 4.1 assumes the human-annotated transition graph is complete and noise-free, yet the paper gives no argument that the ASP-minimal PDDL domain learned from one finite graph is the true domain on larger instan","rationale":"The reader's weakest_assumption identified exactly the same risk: the learned PDDL model is only as good as the completeness and correctness of the human-annotated transition graph, and the zero-shot Hanoi result is one positive test rather than a systematic verification. The central claim that the framework generalizes to larger instances depends on the ASP-minimal domain being identifiable from a finite graph. The paper itself flags the completeness/noise-free assumption in Section 4.1, so this is not an invented concern. The framework is otherwise coherent: the pipeline combines existing ASP-based symbolic learning with diffusion policies, and the qualitative results are promising. However, neither the graph-construction process nor the ASP level is evaluated for robustness, and the empirical section lacks the quantitative detail needed to confirm the strong generalization claim. A CONDITIONAL verdict is appropriate, with the concrete test above being a minimal additional check that would materially strengthen or weaken the central claim.","tokens_in":15566,"tokens_out":7152,"duration_ms":88673,"concrete_test":"Use the published 3x3 Hanoi graph (Fig. 8a) and the learned PDDL domain (Appendix A.1). Generate perturbations that simulate incompleteness/noise: remove 1, 5, or 10 random MOVE edges, and also merge two visually similar states into one node. For each perturbed graph, re-run the ASP abstraction, instantiate the resulting domain on 4x3, 5x4, and 7x5 towers, and check whether the induced ground transition relation equals the true legal-move relation. Also run MetricFF plans through a no-op environment. If any perturbation yields illegal transitions or fails to reach a solvable goal, the completeness/noise assumption is essential and the single clean-graph result is not robust evidence of generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 explicitly states: \"We assume graphs are complete and noise-free.\" This is load-bearing because the ASP solver returns the simplest PDDL domain whose state graph is isomorphic to the annotated graph G, and any finite G is consistent with infinitely many domains that disagree on unseen instances. The paper's only evidence that the minimal model coincides with the true Hanoi theory is the single zero-shot transfer from 3x3 to 7x5 (Fig. 5). If the annotated 3x3 graph omits legal transitions, or if any visual snapshot matching is inconsistent, the learned MOVE operator may be unsound or incomplete when instantiated for 4x3, 5x4, or 7x5 configurations. Because the graph is built through human snapshot matching and label assignment (Section 4.1), completeness is not guaranteed by any automatic procedure, and the claimed generalization to larger task instances is therefore contingent on an unverified identifiability property. The cited robustness of the ASP solver to noise in prior work does not transfer automatically to this pipeline without a concrete demonstration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neuro-symbolic imitation learning framework that learns, from a small number of skill demonstrations, both a PDDL-style symbolic domain and the low-level neural policies needed to execute plans. Demonstrations are first converted into a labeled transition graph via human state matching and skill labels; the graph is minimized by bisimulation and passed to an ASP solver to obtain a minimal symbolic domain. A learned planner produces abstract plans, and each operator is realized by a diffusion policy trained on demonstration segments, with an oracle filtering observations and expressing them relative to the end effector. The authors claim state-of-the-art data efficiency (as few as five demonstrations), zero-shot transfer to larger Hanoi instances and new forklift tasks, and interpretable symbolic plans. The method is evaluated in six simulated domains.","tokens_in":15794,"tokens_out":6076,"duration_ms":72394,"significance":"If the empirical claims are substantiated, the framework is a meaningful step toward joint learning of symbolic abstractions and continuous controllers from few demonstrations, without hand-designed predicates or lexicons. The paper's concrete strengths are the explicit pipeline from raw trajectories to bisimulation-minimized graphs to ASP-discovered PDDL domains, the oracle-based observation filtering, and the interpretable generated domains shown in Appendix A.1. The use of a classical planner and modular skills also gives a credible route to long-horizon generalization. However, the manuscript does not provide the quantitative evidence needed to support its central claims, and the symbolic-generalization argument rests on an identifiability assumption that is stated but not justified. The paper ships no code or machine-checked proofs, so the reported results are not independently reproducible from the text alone.","major_comments":[{"comment":"The central empirical claim is that N-S achieves high data efficiency and strong zero-/few-shot generalization. However, Section 5 states that each agent is evaluated over 30 episodes with 5 seeds, but Section 6 and Figs. 4–5 report no numerical success rates, standard deviations, confidence intervals, or per-seed breakdowns. Phrases such as 'near 100% success' and 'baselines completely fail' are qualitative. Without quantitative tables or error bars, the comparisons and the generalization results cannot be assessed. This is load-bearing because the paper's contribution is primarily empirical.","section":"Section 6, Figs. 4–5"},{"comment":"The comparison to baselines is not apples-to-apples. The N-S method uses a high-level symbolic planner and an oracle that filters observations and provides relative coordinates; the H-IL and IL baselines do not have access to this planner or to the oracle filtering γ. The paragraph on 'fair access' only describes observation-coordinate differences, not the presence of the planner/oracle. Consequently, the reported success of N-S may reflect the privileged planner/oracle rather than the jointly learned symbolic abstraction. To support the claim that both layers are learned, the paper should include an ablation in which H-IL receives the same ground-truth subgoals, or a version of N-S without the symbolic planner, so that each component's contribution can be isolated.","section":"Section 5, Baselines"},{"comment":"The identifiability of the ASP-minimal PDDL domain is assumed but not established. The paragraph states 'We assume graphs are complete and noise-free.' Since the graph is built from human visual snapshot matching and label assignment, completeness is not guaranteed by construction. Any finite labeled graph is consistent with infinitely many PDDL domains that disagree on unseen states; the ASP solver returns the simplest one. The only evidence that the minimal model matches the true Hanoi domain is the single zero-shot transfer in Fig. 5 (3×3 to 7×5), which is presented without error bars. The limitation statement in Section 7.1 invokes prior work [14] on ASP robustness to noise; that robustness concerns a different pipeline and does not cover errors introduced by snapshot matching. The paper should provide an identifiability argument or, at minimum, an additional validation (e.g., held-o","section":"Section 4.1, last paragraph"},{"comment":"The paper frames the input as 'raw skill demonstrations' and says 'Human input is limited to two forms' (Section 3), but the experiments use demonstrations generated by a 'hand-coded automated script' with injected Gaussian noise (Appendix A.5), not human demonstrations. This is a mismatch with the claimed human-taught setting. The manuscript should either clarify that the reported results are on scripted stand-ins and discuss how human demonstrations would affect snapshot matching and graph completeness, or report experiments with actual human demonstrations. This matters because the human annotation step is one of the paper's stated contributions.","section":"Problem Formulation and Appendix A.5"},{"comment":"The statement 'no prior work jointly learns low-level control policies and high-level planning models from a few demonstrations without relying on predefined symbolic states...' is not substantiated relative to the cited reference [50], whose title is 'Neuro-symbolic imitation learning: Discovering symbolic abstractions for skill learning.' Unless the authors show that [50] requires predefined symbolic states or differs in a concrete technical way, the 'first' claim is unverified. Please add an explicit comparison of assumptions, inputs, and mechanisms with [50], or soften the claim.","section":"Introduction, Contributions; Related Work"}],"minor_comments":[{"comment":"There is a typo: 'Nut Assemly' should be 'Nut Assembly'.","section":"Fig. 4 caption"},{"comment":"The phrase '0 secondes' should be '0 seconds'.","section":"Appendix A.5"},{"comment":"The notation Doi is used in 'Cluster Action Steps(πi, Doi)' but is not formally defined before this line. Please define it in Section 4.","section":"Algorithm 1, line 5"},{"comment":"The oracle function ϕ is clear in principle, but the text does not state how the relevant-object set E_oi is computed from the ASP-discovered operator. Is it derived from predicate changes in the operator's effects? Please specify the procedure.","section":"Section 4.2, Eq. (1)–(2)"},{"comment":"The action step clustering threshold is never given a value or a sensitivity analysis. Since it is a free parameter, the paper should report the chosen threshold and, ideally, an ablation.","section":"Section 5"},{"comment":"No code, dataset, or detailed hyperparameter tables beyond Appendix A.5 are provided. A public release of the demonstration-generation scripts, learned PDDL domains, and evaluation harness would be necessary to reproduce the reported results.","section":"General reproducibility"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth your time: build a transition graph from a handful of human-annotated demonstrations, let an ASP solver extract a minimal PDDL domain, then train diffusion policies for each operator with an oracle that filters observations to the relevant objects. That integration is new as far as I know, and the zero-shot transfer from 3x3 to larger Towers of Hanoi configurations is the kind of test that actually exercises the symbolic generalization claim. The paper is also honest about its main assumption -- that the annotated graph is complete and noise-free -- and it points to prior work on ASP robustness rather than pretending the problem doesn't exist.\n\nThe soft spots are real but addressable. First, the evaluation is under-reported: no success rates, no error bars, just 'near 100%' and 'baselines completely fail' in prose, even though the method section says 30 episodes and 5 seeds. That needs tables and variance. Second, the baselines lack both the oracle and the symbolic planner, so it is unsurprising they collapse on long-horizon tasks; the comparison does not isolate where the benefit comes from. Third, the 'as few as five demonstrations' claim is weaker than the text suggests -- the paper's own Fig. 5 says 5-demo performance in 3x3 was limited and improved only after expert corrections. That's fine, but it should be stated honestly in the abstract. Fourth, the symbolic generalization rests on an identifiability assumption: a finite graph can be consistent with many domains, and the single 7x5 success is encouraging but not strong evidence that the learned MOVE operator is correct for all larger instances. The authors acknowledge this in the limitations, but the paper would be much stronger with an explicit test of the learned domain on a held-out configuration that requires a genuinely novel rule application.\n\nThe thinking in the paper is clear and the piece is well situated in the literature. This is not a paradigm shift, but it is a solid integration with a believable mechanism. I would send it to peer review, expecting major revision on the evaluation side. If the quantitative results hold up after the revision, I would cite it.","headline":"A genuinely useful integration of ASP-based symbolic learning with diffusion-policy skills, and the zero-shot Hanoi transfer is the right kind of evidence, but the thin quantitative reporting and the 'five demonstrations' claim make the current version a strong workshop paper rather than a finished contribution.","tokens_in":16298,"tokens_out":1928,"would_cite":true,"duration_ms":24627,"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 a robot can jointly learn continuous control policies and the symbolic planning domain of a task from as few as five raw skill demonstrations, with no predefined states, predicates, or hand-coded domain knowledge.","keywords":["neuro-symbolic imitation learning","task and motion planning","symbolic abstraction","PDDL","answer set programming","diffusion policy","long-horizon manipulation","few-shot learning"],"falsifier":"Take the same Towers of Hanoi demonstrations and deliberately corrupt one state snapshot match (or remove one node from the graph), rerun the full pipeline, and test on the 7x5 configuration: if the planner still solves all instances, the framework tolerates annotation noise; if it fails or produces an invalid plan, the completeness/noise-free assumption is doing the load-bearing work.","tokens_in":15448,"feed_emoji":"🤖","tokens_out":7067,"duration_ms":68078,"temperature":0.7,"pith_summary":"The paper's central claim is that long-horizon robotic tasks can be learned end-to-end from a handful of raw skill demonstrations without any predefined symbolic vocabulary: the system jointly discovers low-level continuous controllers and the high-level PDDL domain needed to plan them. Each demonstration is treated as an edge between two black-box high-level states that a human matches by looking at before/after images, and an Answer Set Programming solver finds the minimal symbolic domain consistent with the resulting graph. A learned oracle filters each controller's observation down to the objects the symbolic operator actually changes, expressed relative to the end-effector, so diffusion policies can be trained from very few trajectories. The authors report that with about five demonstrations per skill the approach solves Robosuite tasks (Stacking, Kitchen, Nut Assembly), a Towers of Hanoi implementation, and forklift pallet tasks, and that the learned Hanoi domain transfers zero-shot from 3x3 to 7x5 instances. The manuscript's Section 7.1 acknowledges that success depends on demonstration quality, accurate grasp-pose estimation, and the oracle preserving the Markov property.","feed_headline":"Five demos yield a plan and the skills to execute it","feed_subtitle":"A visual snapshot graph and an ASP solver replace hand-coded symbols, so long-horizon tasks generalize without retraining.","key_machinery":"The load-bearing object is the labeled transition graph of black-box states: nodes are high-level states identified only by visual snapshots, edges are human-labeled skill transitions. Bisimulation minimization removes redundant nodes, then an ASP solver treats the graph as a specification and searches for the minimal PDDL domain whose state graph is isomorphic to it. Around that core, the oracle phi = alpha composed with gamma filters each skill's observation space to the objects its symbolic operator changes and re-expresses their coordinates relative to the end-effector, and diffusion policies provide the continuous low-level controllers. The PDDL domain connects the two levels: a classic","core_discovery":"The discovery the paper is trying to establish is that symbolic abstraction and neural control can be learned jointly from sparse demonstrations, removing the usual requirement for manually designed predicates, object types, or lexicons. The pipeline maps each raw skill trajectory to a labeled node transition (n, l, n') with visual snapshots; humans only decide whether two snapshots show the same high-level state. A bisimulation quotient compacts the graph, and an ASP solver recovers the simplest first-order planning domain (expressed in PDDL) whose state graph is isomorphic to the observed one. Each discovered operator is then paired with a diffusion policy trained on the corresponding demo","pith_inferences":["Editorial inference: the human visual-matching bottleneck is the main remaining supervision cost; the paper names vision-language models as a possible automation, but the data-efficiency claim depends on how cheap this matching is in practice.","Editorial inference: 'five demonstrations' counts skill demonstrations, not full task traces, so the total human effort for a long-horizon task scales with the number of distinct skills and states the graph needs.","Editorial inference: the ASP solver's minimality criterion is a simplicity prior; if the same transition graph admits two equally simple PDDL theories with different generalization behavior, the method would silently pick the wrong one, so broader zero-shot tests across multiple domains would stress this.","Editorial inference: the oracle assumes the symbolic operator identifies all task-relevant objects; tasks where relevance depends on continuous or unobserved quantities (e.g., force, temperature, occlusion) would need a richer symbolic vocabulary than a sparse graph can invent."],"forward_implications":["Long-horizon manipulation no longer needs hand-coded symbolic models: a few demonstrations plus visual state matching can produce a planning domain and the controllers to execute it.","The learned symbolic abstraction supports zero-shot generalization to larger instances and spatial shifts, since planning happens in compact PDDL space rather than in raw observation space.","Adding new skills, object types, or predicates can be done by extending the graph with a few nodes and demonstrations, enabling curriculum-style fine-tuning.","Because the plans are symbolic, a human can inspect and correct the robot's high-level reasoning before any execution.","The approach is domain-agnostic: the same pipeline works for Cartesian manipulator arms and for an articulated forklift with rear-wheel steering."],"supporting_citations":[{"why":"Supplies the ASP-based method that learns first-order symbolic representations (the PDDL domain) from the black-box state graph.","marker":"[13, 14]"},{"why":"Supplies the diffusion policy model used to train low-level controllers from filtered demonstration segments.","marker":"[55]"},{"why":"Defines the two desiderata (KD1/KD2) for bilevel neuro-symbolic skill learning and represents the closest prior joint-learning approach the paper must exceed.","marker":"[31]"},{"why":"MetricFF is the classical planner used to compute abstract operator sequences from the learned PDDL domain.","marker":"[54]"},{"why":"Baseline that induces symbolic abstractions via clustering and needs fifty or more demonstrations, against which the paper's five-demonstration efficiency is measured.","marker":"[23]"},{"why":"PDDL is the formal language into which the learned domain is translated for classical planning.","marker":"[36]"},{"why":"Options framework inspiration for decomposing each skill into sequential action steps with reduced action spaces.","marker":"[53]"},{"why":"Cited for the ASP solver's robustness to noisy or missing nodes and edges, which the paper invokes in its limitations discussion.","marker":"[14]"}],"fun_headline_variants":["Five demos teach skills and symbolic planning rules","Neuro-symbolic imitation from five demonstrations","Five demos: joint learning of control and symbols","Graph and solver: five demos to plan and act","Five skill demos build generalizable planning and acting"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the human-annotated graph of state transitions is complete and noise-free enough that the ASP solver's minimal symbolic domain is the true generalizable domain; miss a key state or mismatch a snapshot and the learned PDDL model will be wrong for unseen configurations.","fun_headline_variants_meta":{"raw":{"variants":["Five demos teach skills and symbolic planning rules","Neuro-symbolic imitation from five demonstrations","Five demos: joint learning of control and symbols","Graph and solver: five demos to plan and act","Five skill demos build generalizable planning and acting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1789,"prompt_tokens":727,"completion_tokens":1062,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":1002}},"tokens_in":471,"tokens_out":1062,"duration_ms":10892,"temperature":1.0,"reasoning_tokens":1002,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:16:54.862117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same Towers of Hanoi demonstrations and deliberately corrupt one state snapshot match (or remove one node from the graph), rerun the full pipeline, and test on the 7x5 configuration: if the planner still solves all instances, the framework tolerates annotation noise; if it fails or produces an invalid plan, the completeness/noise-free assumption is doing the load-bearing work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion policy model used to train low-level controllers from filtered demonstration segments."},{"cited_title":"Silver, A","cited_arxiv_id":null,"evidence_quote":"Defines the two desiderata (KD1/KD2) for bilevel neuro-symbolic skill learning and represents the closest prior joint-learning approach the paper must exceed."},{"cited_title":"ignoring delete lists","cited_arxiv_id":null,"evidence_quote":"MetricFF is the classical planner used to compute abstract operator sequences from the learned PDDL domain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline that induces symbolic abstractions via clustering and needs fifty or more demonstrations, against which the paper's five-demonstration efficiency is measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PDDL is the formal language into which the learned domain is translated for classical planning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited for the ASP solver's robustness to noisy or missing nodes and edges, which the paper invokes in its limitations discussion."}],"review_version":1}