{"id":"8c43d90f-9012-45c4-b961-a7918f0b3b4e","arxiv_id":"1908.06769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A continuous relaxation of a symbolic planner, operating on probabilistic symbol grounding outputs, lets one-shot imitation learning succeed with far fewer meta-training tasks than policy-network baselines.","lead":"This paper makes one-shot imitation learning more data efficient by splitting it into a learned symbol grounding step and a planning step, and lets the planner work directly on probabilities instead of hard yes/no symbols. A reader interested in robots that learn new tasks from a single demonstration would want to see whether this planning-based decomposition actually removes the need for hundreds of training tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Action updates do not preserve the factorized independence that Eq. (6) relies on; multi-step Continuous Planner is an unstated approximation, not the claimed exact relaxation.","rationale":"The paper's central claim is that replacing the set-theoretic state with probabilistic symbols and planning directly on SGN outputs yields a continuous relaxation that significantly improves one-shot imitation. The reader's weakest assumption was the deterministic, known transition model, which the paper explicitly acknowledges. I find a more internal problem: even under deterministic transitions, the update equations only maintain marginals, but the next-step applicability formula Eq. (6) requires the full joint distribution unless the ground atoms are independent. Deterministic action application creates dependencies among atoms (e.g., an added atom is correlated with the action's precondition), so the factorized form is not preserved. The paper does not flag this approximation anywhere. This is a correctness risk for the theoretical derivation, though the experiments might still show empirical gains if the approximation is mild in the two tested domains. Because the flaw is addressable by either maintaining the exact joint, providing bounds, or empirically comparing against exact belief tracking, the existing CONDITIONAL verdict remains appropriate. I agree with the reader's overall assessment but believe the specific load-bearing assumption is the implicit post-update independence, not stochasticity of the environment.","tokens_in":12074,"tokens_out":13250,"duration_ms":134847,"concrete_test":"Run exact joint inference on a minimal STRIPS domain with initial P(A)=0.5, P(B)=0.1 independent; action o1 precondition A, add B (no delete); action o2 precondition A∧B, add C. Compare CP's computed applicability of o2 (Eq. 6 after Eq. 11) with exact enumeration over the 4 states. If they differ, the iterative marginal update is not an exact relaxation. Also rerun the Block Stacking evaluation with an exact joint-belief version of the Continuous Planner; if success rates match, the approximation is benign for the reported domains, if not, the paper's explanation of the gains is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest point is not stochasticity but a hidden independence assumption in the iterative update. The paper represents Z(s) by marginals under conditional independence, and Eq. (6) computes an action's applicability as the product of the marginal probabilities of its preconditions. Eqs. (10)-(15) then update these marginals after attempting an action. But even with deterministic, known operators, the post-action joint distribution is generally not independent: an added atom B after an action with precondition A satisfies B' = A∨B, so B' is correlated with A. After the first update, a later action whose precondition is A∧B has true applicability P(A), while the factorized formula P(A)P(B') can be much smaller. The paper does not state or justify an assumption that the distribution remains factorized after each action; it simply continues to use only marginals. Thus multi-step plans are computed from an approximation, not the claimed relaxation. This is testable and is not covered by the acknowledged deterministic-transition assumption. Example: initial P(A)=0.5, P(B)=0.1, action o1 Pre={A}, Eff={B}; after Eq. (11), P(B')=0.55; for action o2 Pre={A,B}, Eq. (6) gives 0.275, while exact enumeration gives P(A)=0.5. This invalidates the derivation for arbitrary STRIPS domains unless additional domain-specific constraints rule out such dependencies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for one-shot imitation learning that decomposes the problem into symbol grounding and planning. A modular Symbol Grounding Network (SGN) maps continuous states to probabilistic symbolic states, and a Continuous Planner (CP) performs planning directly on these probabilistic states instead of discretizing them into a single symbolic state. The authors derive iterative update formulas for action applicability, action effects, and goal satisfaction under the assumption of deterministic symbolic transitions. They evaluate the approach on Block Stacking and Object Sorting domains, reporting that the planning-based methods, especially the Continuous Planner, outperform the Neural Task Graph Networks baseline with fewer meta-training tasks, and that the continuous relaxation offers gains over a symbolic planner that operates on discretized SGN outputs.","tokens_in":12328,"tokens_out":2985,"duration_ms":32892,"significance":"If the central claims hold, the paper makes a useful contribution by demonstrating that explicit symbolic planning, when combined with a learned probabilistic grounding, can reduce the meta-training burden in one-shot imitation learning. The proposed modular SGN and the idea of planning on continuous symbol probabilities are reasonable and interesting extensions of prior work, and the empirical results on two domains suggest that the approach is promising. The formulation disentangles policy execution from inter-task generalization, and the method does not rely on handcrafted rules for invalid states. However, the derivation of the continuous planner rests on an unstated independence assumption that is not preserved by the updates, and the goal-matching objective is left unspecified; these issues affect the paper's main technical contribution and must be addressed before the claims are fully supported.","major_comments":[{"comment":"The derivation of the iterative update formulas assumes that the distribution over symbolic states remains factorized after each action update, but this is not stated and generally is false. For example, consider an operator o1 with precondition {A} and effect {B}; after applying the update in Eq. (11), the marginal P(B') becomes P(A) + (1-P(A))P(B), but the joint distribution has P(B'|A)=1, so B' is positively correlated with A. Consequently, for a subsequent operator o2 with precondition {A,B}, the true applicability probability is P(A), while Eq. (6) would return P(A)P(B'), which can be arbitrarily smaller. The paper's claim that the formulas are exact and avoid marginalizing a large state space (Section IV-C, after Eq. (15)) therefore only holds under a mean-field approximation that is not acknowledged. Please state this assumption explicitly, justify it empirically or theoretically, or restrict the claims to cases where the domain structure preserves conditional independence.","section":"Section IV-C, Eqs. (7)-(15)"},{"comment":"The goal-satisfaction step is described only as \"the objective of our search is thus to match the two distributions,\" but no concrete objective or algorithm is given. Without a definition of the distance or divergence between the current state distribution and the goal distribution, it is impossible to reproduce the planner's search procedure or to verify that the planner is actually optimizing the stated criterion. Please provide a formal definition (e.g., cross-entropy, KL divergence, or a weighted sum of atom probabilities) and explain how it is used in the forward search.","section":"Section IV-C, \"Goal Satisfaction\""},{"comment":"The success-rate curves are reported without error bars, the number of random seeds, or any statistical significance testing. The central claim that the Continuous Planner \"significantly outperforms\" the symbolic planner (Section V-B) may be sensitive to random initialization of the SGN and the variability of the evaluation tasks. Please report means and variances over multiple seeds and state the number of seeds used, or otherwise provide a statistical basis for the comparison.","section":"Section V, Figures 5 and 6"}],"minor_comments":[{"comment":"There is a typo in the caption: \"Symbol Gnding Networks\" should be \"Symbol Grounding Networks.\"","section":"Figure 1 caption"},{"comment":"In the line \"sc←ExecActions(E, Π)\", the variable sc is reused for the current continuous state after execution, which is also the name of the initial continuous state s0; using a distinct name such as s_{next} would improve clarity.","section":"Algorithm 2"},{"comment":"The notation Π = CP(SGN(sc), SGN(dτ_T), O) is introduced but the closed-loop nature is only described in words; it would be clearer to specify how the plan is executed and re-planning is triggered, since Algorithm 2 leaves the exact interaction with the environment implicit.","section":"Section IV-A, Eq. (1)"},{"comment":"The sentence \"SymbolicState(·) computes the aligned symbolic state Sτ for each dτ based on the action annotation aτ used in previous works\" relies on references [4,6] for the action annotations; please clarify whether these annotations are available at meta-training time for all tasks and how they are obtained for novel tasks at test time.","section":"Section IV-D, Learning"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the high-level idea is attractive, but the technical derivation contains an unacknowledged factorization assumption that could undermine the exactness claim of the continuous planner. The reviewer's specific counterexample is not covered by the acknowledged deterministic-transition assumption and should be addressed with either a formal statement of the mean-field approximation or a domain-specific argument. The lack of any statistical reporting in the experiments is also concerning for a paper that advertises a significant improvement. The contribution is potentially salvageable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the bottom line: the paper is worth a serious look. Its central empirical claim—that planning on probabilistic symbols from a learned grounding network works much better than discretizing those symbols—is supported by the experiments. The modular Symbol Grounding Network is a real contribution, and the framing of one-shot imitation as planning rather than end-to-end policy learning is a sensible way to reduce meta-training data.\n\nWhat's actually new: the integration of a learned grounding network with a planner that operates on the distribution over symbolic states, plus the modular parameter sharing for grounding. The baselines are reasonable, and the comparison to NTG makes the case that the disentanglement helps.\n\nThe soft spots, in rough order of importance. First, the stress-test note is right: the iterative update in Eqs. (10)–(15) silently assumes that the distribution over symbolic states stays factorized after each action. Even with deterministic operators this is false. If action o1 has precondition A and adds B, then B' is A∨B, so B' is correlated with A. A later action with precondition A∧B then has true applicability P(A), but the factorized formula gives P(A)P(B'), which can be much smaller. The paper calls this a continuous relaxation of the symbolic planner, but multi-step planning is using an approximation without saying so. That is a genuine gap in the derivation, and it should be either fixed or explicitly acknowledged. The experiments may still make the method useful, but the claim of deriving an exact relaxation is too strong.\n\nSecond, the goal-matching objective in Section IV-C is never defined. The planner is told to 'match the two distributions', but there is no equation for the search objective. That makes the planner hard to reproduce. Third, the success-rate curves have no error bars, seed counts, or confidence intervals; given that the reader's main conclusion rests on the gap between CP and SP, this is a minor but real weakness.\n\nThe citation pattern is fine; Konidaris et al. is credited for the probabilistic symbol representation, and the self-citations for action annotations are legitimate. No circularity.\n\nWho is this for? Anyone working on combining learning with symbolic planning, and the empirical setup is a useful benchmark. It deserves a serious referee: the problem is important, the empirical results are promising, but the paper needs to either fix the independence approximation, define the planner objective, and add statistical rigor. I would not desk-reject it; I would send it back with these requests.","headline":"The empirical claim about planning on probabilistic symbols is plausible and worth testing, but the planner derivation hides an independence approximation that makes multi-step planning inexact; the paper should acknowledge this and define its goal-matching objective.","tokens_in":12849,"tokens_out":3165,"would_cite":true,"duration_ms":30023,"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":"This paper argues that one-shot imitation learning can be recast as symbolic planning over probabilistic symbol groundings, so a learned network's uncertain outputs feed a continuous planner directly instead of being forced into discrete…","keywords":["one-shot imitation learning","symbolic planning","continuous relaxation","probabilistic symbols","symbol grounding","modular neural networks","data efficiency","long-horizon manipulation"],"falsifier":"Run the Continuous Planner in a domain where one action succeeds with probability 0.8 and otherwise leaves the state unchanged, and compare the planner's predicted distribution after one action (Eqs. 10-15) with the empirical distribution over states. If they diverge, the deterministic-transition update is invalid. Likewise, a domain with two strongly anti-correlated ground atoms (e.g., an object cannot be in two containers at once) should expose the conditional-independence assumption as the source of planning error.","tokens_in":11904,"feed_emoji":"🤖","tokens_out":7379,"duration_ms":62576,"temperature":0.7,"pith_summary":"This paper tries to show that one-shot imitation learning—executing a never-seen task from a single demonstration—can be made far more data-efficient by separating the job into two pieces: a learned module that turns continuous observations into symbolic facts, and a planner that executes the task from those facts. The authors' key move is to let the learned module output probabilities instead of hard symbols, and to relax the symbolic planner so it plans directly on those probabilities rather than forcing a discrete, often invalid symbolic state. On two simulated manipulation domains, the resulting Continuous Planner succeeds with only 8 to 15 meta-training tasks, where a leading neural approach needs hundreds. A sympathetic reader would take the paper's claim to be that this decoupling, plus the continuous relaxation, is what buys the data efficiency.","feed_headline":"Planning on fuzzy symbols lets one-shot imitation learn from 8 tasks","feed_subtitle":"Discretizing the network's guesses breaks the planner; planning on its probabilities matches models trained on hundreds of tasks.","key_machinery":"The load-bearing object is the probabilistic symbol representation from [10], which replaces a set of true ground atoms with a distribution over symbolic states, compactly encoded by the marginal probability per ground atom. On top of it, the Continuous Planner derives iterative formulas for state update, action applicability, and goal matching that operate only on those marginals, using the assumption that ground atoms are conditionally independent. The modular Symbol Grounding Network, with object modules and predicate modules that share parameters, supplies those marginals and is the only learned component.","core_discovery":"The central claim is that the compound mapping from demonstration to policy, which prior one-shot methods learn as a single black box, can be decomposed as $\\phi(\\cdot)=CP(SGN(\\cdot))$: a Symbol Grounding Network maps continuous states to distributions over ground atoms, and a Continuous Planner searches for an action sequence that takes the current state distribution to the goal distribution. Because the planner assumes a known deterministic transition model, the effect of attempting an action on each ground atom's probability can be written in closed form (Eqs. 10-15) under a conditional-independence assumption, so no large state-space marginalization is needed. This makes it possible to plan on uncertain symbolic input directly, and the planner reduces to the classical symbolic planner when the distribution concentrates on one state. The paper reports that this formulation beats the symbolic-planner baseline that discretizes the same network outputs, and matches a manually-heuristic baseline without needing domain-specific rules.","pith_inferences":["The same continuous relaxation should transfer to any classical planner with a known deterministic transition model, so task-and-motion-planning systems that currently threshold learned classifier outputs could adopt it directly; the expected benefit is the same robustness to grounding errors.","The conditional-independence assumption on ground atoms is the main simplification; a domain with strong correlations between facts (e.g., spatial exclusivity) might need a richer state distribution than per-atom marginals. Modeling those correlations would be a natural next step.","A strong test of the paper's decoupling thesis is to replace the modular SGN with a non-modular network of the same capacity; if performance degrades, that would confirm that parameter sharing among symbols, not just the planner relaxation, drives the data efficiency.","The planner in Eq. (1) only sees the final demonstration state unless the goal recognizer is conditioned on the full demonstration; the closed-loop formulation is likely to be more robust when the final frame is ambiguous. This is an extension the paper mentions but does not evaluate."],"forward_implications":["On Block Stacking, planning-based methods all outperform the policy-network baseline (NTG) at every training-set size, and the Continuous Planner is the strongest, even matching a manually engineered heuristic without hand-defined rules.","On Object Sorting, the Continuous Planner reaches 100% success with only 8 training tasks, a regime where neither NTG nor the symbolic planner converges within 15 tasks.","Because the Continuous Planner is a generalization of the symbolic planner, it inherits the ability to find alternative solutions to the same goal, which the paper shows is essential for the Object Sorting task.","Only the Symbol Grounding Network needs training; the planner itself is derived, so the approach's data requirement is set by how well the SGN can learn to ground symbols, not by how well a policy network can imitate."],"supporting_citations":[{"why":"Supplies the probabilistic symbol representation that the continuous relaxation is built on.","marker":"[10]"},{"why":"Defines the classical set-theoretic symbolic planning formalism that the paper extends.","marker":"[9]"},{"why":"Provides the Neural Task Graph baseline and the Block Stacking / Object Sorting task setups.","marker":"[4]"},{"why":"Supplies the Neural Task Programming baseline and the prior data-hungry one-shot imitation setup.","marker":"[6]"},{"why":"Gives the DET-POMDP formulation whose deterministic-transition assumption the paper borrows and relaxes.","marker":"[14]"}],"fun_headline_variants":["Fuzzy-symbol planning boosts one-shot imitation from just 8 tasks","Probabilistic planner beats discretization for one-shot imitation","Continuous planning makes one-shot imitation work with just 8 tasks","Don't discretize: plan on probabilities for one-shot imitation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-update formulas assume the task domain's transition rules are known and deterministic, and that the network's fact probabilities are independent of one another; if actions can fail or facts are correlated, the planner's predicted states may be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Fuzzy-symbol planning boosts one-shot imitation from just 8 tasks","Probabilistic planner beats discretization for one-shot imitation","Continuous planning makes one-shot imitation work with just 8 tasks","Don't discretize: plan on probabilities for one-shot imitation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2326,"prompt_tokens":891,"completion_tokens":1435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":1363}},"tokens_in":507,"tokens_out":1435,"duration_ms":10720,"temperature":1.0,"reasoning_tokens":1363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:57:29.436500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Continuous Planner in a domain where one action succeeds with probability 0.8 and otherwise leaves the state unchanged, and compare the planner's predicted distribution after one action (Eqs. 10-15) with the empirical distribution over states. If they diverge, the deterministic-transition update is invalid. Likewise, a domain with two strongly anti-correlated ground atoms (e.g., an object cannot be in two containers at once) should expose the conditional-independence assumption as the source of planning error.","supporting_citations":[{"cited_title":"Symbol acquisi- tion for probabilistic high-level planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic symbol representation that the continuous relaxation is built on."},{"cited_title":"Ghallab, D","cited_arxiv_id":null,"evidence_quote":"Defines the classical set-theoretic symbolic planning formalism that the paper extends."},{"cited_title":"Neural task graphs: Generalizing to unseen tasks from a single video demonstration,","cited_arxiv_id":null,"evidence_quote":"Provides the Neural Task Graph baseline and the Block Stacking / Object Sorting task setups."},{"cited_title":"Neural task programming: Learning to generalize across hierarchical tasks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Neural Task Programming baseline and the prior data-hungry one-shot imitation setup."},{"cited_title":"Algorithms for sequential decision making,","cited_arxiv_id":null,"evidence_quote":"Gives the DET-POMDP formulation whose deterministic-transition assumption the paper borrows and relaxes."}],"review_version":1}