{"id":"63e0eae4-d4f2-4aed-9b7e-030ce42293b9","arxiv_id":"2412.11829","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Policies retrieved by optimizing a parameter-weighted advantage function outperform policies conditioned on a single estimated parameter in contact-rich manipulation, given a roughly known parameter distribution.","lead":"Robotic manipulation skills can be made robust without knowing the exact mass or friction of an object if the robot averages its action choices over a range of plausible parameters. This paper describes such a method, tests it on hitting, pushing, and reorienting, and reports lower final errors than policies that use a single imperfect parameter estimate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Proposition 2 is invalid: Eq. (15) equates the distributional value with the expectation of parameter-specific oracle values, which is false for closed-loop policies; a two-step counterexample shows the domain-contraction argmax can be strictly suboptimal.","rationale":"Good-faith reading: the paper proposes a useful mechanism for retrieving a policy from a parameter distribution using tensor trains, and the TT efficiency claims in Tables 2 and 3 and the sensitivity analysis in Fig. 15 are credible and independently testable. The real-robot demo is qualitative but shows reactivity. The load-bearing problem is the theoretical superiority over EMA. Prop. 2's proof is not merely missing a step; the claimed optimality is false for multi-step closed-loop tasks because Eq. (15) identifies a distributional value with an expectation of per-parameter oracle values. The constructed counterexample uses the paper's own objective in Eq. (24) and shows the averaged-advantage policy can be strictly worse than an alternative single policy. This is an internal inconsistency, not an issue of outside consensus. The reader's bandwidth concern is real but secondary: it concerns calibrating P, whereas this concern holds even with a perfectly calibrated P. I therefore partially agree with the reader: their rationale notes that the proof is definitional, but their selected weakest assumption (bandwidth) is not the most load-bearing. Because the central theorem is false as stated, the paper should not be accepted on its current theoretical footing. A revised version that reframes domain contraction as a heuristic supported by empirical comparisons, or that learns the belief-state value function directly, could be reconsidered.","tokens_in":21870,"tokens_out":12845,"duration_ms":129149,"concrete_test":"Implement the two-parameter, two-step MDP described above and compare two policy computations. (i) Domain contraction: compute parameter-specific advantage functions with α-oracle continuation values, average them with P = (0.5, 0.5), and take argmax at each state. (ii) Belief-MDP DP: set V_P(s1) = max_v E_α[R_α(s1,v)] = 55, V_P(s2) = 60, and V_P(s0) = max(V_P(s1), V_P(s2)) = 60. Check whether the Eq. (13) policy selects action a and achieves expected return 55 < 60; if so, Proposition 2 is falsified. To connect to the paper's tasks, repeat the comparison on the Push or Reorientation dynamics with two candidate parameter values (e.g., two friction coefficients), comparing closed-loop returns under the averaged-advantage policy and the belief-MDP optimal policy.","verdict_should_be":"REJECT","load_bearing_attack":"The central theoretical claim (Prop. 2, Eq. 27) rests on Eq. (15), which sets V(x|P(α̂)) = Σ_j P_j V(α̂_j, x). This is not the value function of the deployment problem in Eq. (7): in deployment a single policy must act without knowing α, and future decisions also cannot branch by α. The correct dynamic-programming value is V_P(x) = max_u E_α[R(x,u) + γ V_P(f_α(x,u))], not E_α[A_α(x,u)] with A_α defined through the α-oracle continuation V_α. Using V_α overestimates the return of any shared policy, so the argmax of the averaged advantage is not generally optimal. Counterexample: α1 and α2 are equiprobable, γ=1. From s0, action a leads to s1 for both α; action b leads to s2 for both. At s1, v1 gives reward 100 for α1 and 0 for α2; v2 gives 0 for α1 and 110 for α2. At s2, any action gives 60 for both. Domain contraction evaluates a as (100+110)/2 = 105 and b as 60, so it selects a; but any single policy reaching s1 obtains at most max(50,55) = 55, while b yields 60. Thus Eq. (13) is suboptimal for exactly the objective in Eq. (24), and inequality (27) is false in general. This is an internal inconsistency in the optimality argument, not a disagreement with a consensus baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes implicit motor adaptation (IMA), a framework for contact-rich manipulation under uncertain physical parameters. A parameter-augmented base policy is learned offline in tensor-train (TT) form as an implicit representation of the advantage function. At deployment, a probabilistic estimate of the system parameters (a uniform distribution with mean from an MLP and bandwidth w) is used to retrieve a parameter-conditioned policy via \"domain contraction,\" i.e., by taking the argmax over the probability-weighted sum of parameter-specific advantage functions. The authors claim that IMA is theoretically superior to explicit motor adaptation (EMA), which feeds a single point estimate into the base policy. The paper reports simulations on Hit, Push, and Reorientation primitives and a real-robot planar pushing experiment, showing that IMA outperforms EMA and domain randomization, with efficient core-level TT operations for policy retrieval.","tokens_in":22210,"tokens_out":6017,"duration_ms":56763,"significance":"If the practical claims hold, the paper offers a useful engineering contribution: IMA avoids precise system identification and student-policy retraining, and the TT representation makes probabilistic policy retrieval fast enough for real-time control. The simulation protocol includes multiple randomized instances with error bars, and the real-robot demonstration with disturbances is a valuable stress test. However, the central theoretical proposition on which the paper's optimality claims rest is not correct as stated. The empirical comparison to EMA is still informative and appears to support the weaker, practical claim that IMA works well with rough estimates in these tasks, but the theoretical superiority argument in Proposition 2 is unsound and needs to be replaced or substantially reframed.","major_comments":[{"comment":"Equation (15) defines V(x|P(α̂)) as the expectation of the parameter-specific value functions V(x|α̂_j). This is not the value function of the deployment problem in Eq. (24): a single policy π must act without knowing α, both at the current state and at all future states, so its value under the distribution satisfies V_P^π(x)=E_α[R(x,π(x))+γ V_P^π(f_α(x,π(x)))], not Σ_j P_j V_{α_j}(x). Consequently, the representation A(x,u|P)=Σ_j P_j A_{α_j}(x,u) in Eq. (13) does not follow from Eqs. (14)–(18), and the argmax in Eq. (19) is not generally optimal for Eq. (24). A simple counterexample shows the failure: with equiprobable α1 and α2 and γ=1, from state x0 action a leads to x1 for both parameters while action b leads to x2 for both; at x1 one action gives reward 100 under α1 and 0 under α2, and another gives 0 under α1 and 110 under α2; at x2 any action gives 60. Domain contraction evaluates a as (100+110)/2=105 and b as 60, so it selects a, but any single deployment policy that reaches x1 must choose one action for both parameters, giving at most max(50,55)=55, while b yields 60. Thus the proposed contraction is not the solution of the stated robust-control objective.","section":"§5.3, Eq. (15)"},{"comment":"The inequality J_πi(x,P(α̂)) ≥ J_πe(x,P(α̂)) is asserted in Eq. (27) as a consequence of the previous definitions, but it is not proven. Since the IMA objective in Eq. (26) relies on the same incorrect V(x|P(α̂)) definition identified above, the comparison to EMA does not establish optimality. Moreover, the statement that \"α̂ can be considered a single sample from P(α̂)\" does not by itself imply that the EMA policy's value is bounded by the IMA policy's value; such an inequality requires additional assumptions on the value function (e.g., concavity or smoothness in α) and on the closed-loop behavior of the shared base policy. The empirical results in Section 6.3 are an independent and useful evaluation of the practical claim, but Proposition 2 as stated should be withdrawn or replaced by a proof under explicit and testable assumptions.","section":"§5.4, Eq. (27)"},{"comment":"The bandwidth w of the uniform estimator is a free hyperparameter that is selected after inspecting estimator errors: w=N/20 for Hit, w=N/5 for Push and Reorientation, and Fig. 10 indicates that w=N/3 gives the best results under the achieved estimation accuracy. The robustness claim that IMA works with rough estimates depends on the assumption that the true parameter lies inside the support U(ν_t−w/2, ν_t+w/2); because w is tuned using knowledge of the estimator's error on the evaluation tasks, the experiments do not demonstrate performance under a fixed, a priori uncertainty model. Please report results on held-out instances with w fixed before observing estimator errors, or provide a principled rule for setting w from the estimator's noise covariance rather than from the same test data used for evaluation.","section":"§6.2, §6.3, Fig. 10"}],"minor_comments":[{"comment":"In Eq. (9), the policy in the Bellman operator is written as π_k(x), but the surrounding text and Eq. (8) define the policy as a function of both α and x; the notation should be π_k(α,x) for consistency.","section":"§5.1, Eq. (9)"},{"comment":"The MLP in the probabilistic adaptation module outputs a point estimate ν_t, while Eq. (12) writes a distribution P(α̂_t)=ϕ(·) directly; please clarify that the distribution is assumed uniform with mean ν_t and fixed width w, and state how w is chosen for the estimator.","section":"§5.2"},{"comment":"The factorization P(j1,...,jd)=P_1(α̂_j1)···P_d(α̂_jd) assumes independence across parameter dimensions; if the estimator produces correlated parameter beliefs, the core-level contraction in Eq. (23) no longer applies, so this assumption should be stated explicitly.","section":"§5.3, Eqs. (22)–(23)"},{"comment":"The table header appears to repeat \"Task\" and the reported error/time values are not aligned with the column names as described in the text; please reformat the table and add clear units and definitions of the error metric.","section":"§6.5, Table 3"},{"comment":"There are minor typographical errors, including \"emplyed\" in Section 2.2 and \"arbitary\" in Section 5.3, and the MSE loss in Section 5.2 should be written with a squared norm, i.e., MSE(ν_t, α_t)=∥ν_t−α_t∥².","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is substantial and supports a weaker version of the paper's claim: using a rough probabilistic estimate with domain contraction is a practical and effective alternative to point-estimate EMA for the tested contact-rich primitives. The main obstacle is the theoretical narrative: Proposition 2 and the optimality framing of Eq. (13) are not correct for closed-loop policies. I would not reject the paper because the empirical contributions are valuable and the theory could be reframed as a heuristic argument with supporting experiments. If the authors instead insist on the current theoretical claims without fixing the proof, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xue et al. propose implicit motor adaptation (IMA), which retrieves a parameter-conditioned policy from a probabilistic parameter estimate by contracting a tensor-train advantage function over a distribution. The integration is new relative to their earlier domain-contraction paper, and the simulation comparisons are the real contribution: on Hit, Push, and Reorientation, IMA beats explicit motor adaptation (EMA) when the parameter estimate is rough, and the TT core-level computation is convincingly faster than function-level alternatives. The paper is honest about its building blocks (Xue et al. 2024a, Shetty et al. 2024b), and the empirical setup with ten instances and error bars is adequate for a systems paper.\n\nThe soft spot is Proposition 2, and it is not minor. Equation (15) sets the distributional value V(x|P(α̂)) to the expectation of parameter-specific value functions Σ_j P_j V(α_j, x). That identity only holds if the policy can observe α at every decision point. In deployment, a single policy is executed without knowing α, and future decisions cannot branch by α. The correct value is V_P(x) = max_u E_α[R + γ V_P(f_α(x,u))], not the averaged advantage. A simple two-state counterexample shows that the argmax of Σ_j P_j A_{α_j}(x,u) can be strictly suboptimal for the actual objective in Eq. (24). So inequality (27) is false in general. The authors should either fix the proof or reframe the theoretical claim as a heuristic with empirical support.\n\nThe bandwidth w is also tuned per task, with values like N/20 for Hit and N/5 for Push chosen after inspecting estimator error. That makes the 'no precise estimation needed' claim weaker than advertised. The real-robot section is a demonstration (keyframes and video) rather than a quantitative evaluation, and no code or data are released, which limits reproducibility.\n\nOverall: the empirical pipeline is worth a serious referee. The idea of retrieving policies from distributions over a TT advantage function is useful, and the efficiency numbers are credible. But the paper overclaims in its theory, and a reviewer should push for a corrected version. I would send it to peer review, but recommend major revision: remove or repair the proof, add quantitative real-robot results or clearly label them as qualitative, and make the bandwidth selection principled.","headline":"Solid empirical system paper, but the central theoretical claim does not hold; reviewers should focus on the experiments and demand a corrected or dropped proof.","tokens_in":22720,"tokens_out":4922,"would_cite":true,"duration_ms":41494,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-11T14:32:20.439642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}