{"id":"49d28499-a3be-4771-bad6-e4c3814b0dd2","arxiv_id":"2412.00086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A robot learns dynamic tray carrying from a small set of demonstrations by using an ensemble of learned value functions inside a conservative model-predictive controller.","lead":"This paper shows a robot arm can learn to carry objects on a tray, without gripping them, from only 50 to 100 recorded demonstrations and a controller that plans smooth, safe motions. The method combines learned value functions with model-predictive control and works on a real Franka robot, including objects it never saw during training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generalization claim hinges on Eq. (1)'s assumption that end-effector state alone predicts slip; the controlled real-world experiment holds friction fixed, so cross-friction generalization remains untested.","rationale":"The reader's conditionality is appropriate. The paper has strong empirical support: 800+ real-world trials, ablations for lambda/K, observation spaces, and pessimism schemes. I do not see an internal inconsistency in the derivations; Eq. (2) is a soft-max over ensemble returns, not a formal disagreement penalty, but it is used as a heuristic cost and evaluated empirically. The main load-bearing assumption is indeed the sufficiency of the end-effector observation for predicting slip on unseen objects. My read is partial agreement with the reader: I agree the assumption is weak, but I would not say the value functions 'will not transfer' -- Case Study 2 provides suggestive evidence that some transfer occurs. The problem is the evidence is confounded: training on five objects and using a lenient slip threshold plus conservatism could produce high success without the value function encoding any object-general slip model. The controlled experiment (Case Study 1) deliberately holds friction constant, so it cannot settle the issue. The proposed test isolates friction, the single most object-specific parameter, while holding geometry and inertia fixed. Depending on the outcome, the generalization claim either stands or needs to be narrowed to 'same-friction objects' or 'objects within the training distribution.' Because the paper's current verdict is already conditional, my stress-test does not move it.","tokens_in":14671,"tokens_out":7351,"duration_ms":75829,"concrete_test":"Run 60 trials per condition in simulation: train exactly as in Case Study 1 (single cube, mu=0.3, 50 demos), then evaluate on a cube with identical geometry, mass, and CoM but with tray/object friction changed to mu=0.15 and mu=0.45, using the same goal distribution and success metric. If success at mu=0.15 falls sharply (e.g., below 50%) while mu=0.3 stays near 95%, the value function has not transferred across friction and the household-object successes are likely due to conservatism plus training on multiple objects; if success stays high across all mu values, the reader's sufficiency concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim -- robust transfer to novel objects from 50-100 end-effector-only demonstrations -- depends on Eq. (1) treating x_r_t as a sufficient Markov state for future slip. The contact model in Sec. VI-B contradicts this: the friction-cone constraint (5) depends on object mass, inertia, CoM offset, and friction coefficient through w_GI in Eq. (3). Two objects at the same end-effector state can have different slip margins, so a value function trained on one object's demonstrations has no way to represent that difference. The controlled real-world Case Study 1 does not stress this: it keeps mu=0.3 fixed across training and test objects. Case Study 2 does vary materials, but it trains on 100 demonstrations from five objects and uses a 2cm slip threshold, so its success can be explained by the ensemble soft-max cost (Eq. 2) producing a conservative, object-agnostic policy rather than by value functions that transfer slip knowledge. The observation ablation in Table III reinforces the fragility: including end-effector position actually hurts generalization when start positions change, so the choice of observation space is doing real work and the sufficiency claim is not robust.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Conservative Value MPC (CV-MPC), a method for the 'robot waiter' non-prehensile transport task that trains an ensemble of state-value functions from 50-100 task-space demonstrations labeled only with sparse slip/no-slip costs, then combines the ensemble's return estimates pessimistically via a log-sum-exp objective inside a sampling-based MPC (STORM). The authors evaluate the method in simulation and on a Franka Panda, reporting high success rates on novel convex and household objects, ablations over ensemble size, pessimism temperature, observation spaces, and pessimism schemes, and a simulation study showing improvement over a deliberately biased-friction demonstrator.","tokens_in":1656,"tokens_out":1566,"duration_ms":73070,"significance":"If the claims hold, the paper is practically significant: it demonstrates that a dynamic, contact-rich manipulation skill can be learned from a modest number of end-effector-only demonstrations without explicit object models, and it integrates cleanly with off-the-shelf MPC. The empirical effort is substantial, with over 800 real-world trials across 13 objects, and the paper includes useful ablations, especially the observation-space study (Table III), the initial-state versus pointwise pessimism comparison (Fig. 8), and the biased-friction demonstrator experiment (Fig. 4). The main weaknesses are that the central generalization claim rests on an undefended state-sufficiency assumption, the real-world observation representation is not fully pinned down, and success rates are reported without confidence intervals or significance tests, making several comparative claims hard to evaluate.","major_comments":[{"comment":"","section":"Sec. II / Eq. (1) and Sec. VI-B"}],"minor_comments":[{"comment":"The notation for the pessimistic return is inconsistent: the definition above Eq. (2) uses h in H and u in H, while the equation writes the log-sum-exp over ensemble members, but the relationship between per-member return, the running-cost sum, and the terminal value is not made explicit. Please clarify whether the log-sum-exp is over per-step value sums or terminal values, and make the role of lambda explicit.","section":"Eq. (2)"},{"comment":"There is a typo in '60 trails' (should be '60 trials') in the description of Case Study 2, and the total trial count is described as 'over 800 trials' while the reported case studies sum to 940 real-world trials; please reconcile the count.","section":"Sec. IV-B"},{"comment":"The observation ablation is informative, but the four observation sets should be described in a single consistent notation; currently 'ofull_t', 'xfull_t', and 'wT_ee' are used interchangeably, which makes it hard to map the ablation back to the deployed system.","section":"Sec. VI-C1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution to non-prehensile manipulation and offline RL with MPC, but the two main issues that block acceptance are the unaddressed state-sufficiency assumption behind the generalization claim and the lack of statistical rigor in the success-rate reporting. The absence of real-world baselines is also a concern for a journal-level claim. I would encourage the editor to ask for the changes to be made in a revision rather than rejecting, as the core idea is plausible and the experiments are extensive. It would also help if the authors released the demonstration datasets and code to make the 800+ trials reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real empirical contribution, and the core result—50-100 end-effector demonstrations with sparse slip labels are enough to run a conservative MPC that beats a biased-friction demonstrator on a real Franka—holds up. I don't see a load-bearing error in the math or a circular evaluation. The improvement-over-suboptimal claim is legitimate because the slip labels are true even though the demonstrations come from a biased expert.\n\nWhat's new is the specific combination: ensemble value functions trained from observation-only task-space data, initial-state log-sum-exp pessimism inside a sampling MPC, and hardware evaluation with 800+ trials. The ingredients are known, but the package and the reduced demonstration burden are the point, and the paper makes that case. The ablations are useful, especially the observation-space study showing rotation is the critical input and position can hurt when start positions change.\n\nSoft spots, in proportion. First, the generalization claim is oversold. Case Study 1 keeps the friction coefficient fixed, so it tests shape and inertia generalization within one contact regime, not the 'novel objects' with different materials that the abstract implies. Case Study 2 does vary materials, but it trains on 100 demonstrations from five of the hardest objects before testing on household objects; that is fine, but it is not zero-shot from a single object. Second, Eq. (1)'s sufficiency assumption—end-effector state alone predicts future slip—is doing real work, and the paper never directly tests it against slip variation due to friction or mass distribution. The success in Case Study 2 could come from the ensemble soft-max producing a conservative, object-agnostic policy rather than from transferable slip knowledge. That does not sink the paper, but the authors should say it. Third, reproducibility: no code, no data, and gamma, H, and c_friction are unstated. Those matter for a method paper. Success rates also lack confidence intervals; with 60-80 trials per object they are trivial to compute and would strengthen the claims.\n\nThe math is fine: Eq. (2) is a heuristic penalty, not a formal bound, and the paper uses it as a cost term, which is acceptable. The citation pattern is honest and closest prior work is acknowledged.\n\nWho this is for: anyone working on MPC with learned costs, offline RL for manipulation, or non-prehensile transport. It deserves a serious referee. I would accept it for review conditional on the authors releasing code and parameter settings and tightening the generalization language. If they do that, it is a solid conference paper.","headline":"A solid empirical paper whose core claim holds up, but the novel-object generalization is oversold and the missing code/constants keep it short of an accept as-is.","tokens_in":15437,"tokens_out":1686,"would_cite":true,"duration_ms":16377,"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 dynamic non-prehensile transport can be learned from 50–100 task-space demonstrations with sparse slip labels, and deployed safely by combining an ensemble of value functions with uncertainty-aware model-predictive…","keywords":["non-prehensile manipulation","robot waiter task","model-predictive control","offline reinforcement learning","value function ensemble","sparse reward labels","learning from demonstrations","slip avoidance"],"falsifier":"Train a value-function ensemble on a cube with a fixed surface friction, then test on objects that produce identical end-effector trajectories and nominal friction but differ in internal mass distribution or contact-patch roughness; if success rates collapse while end-effector kinematics are held fixed, the Markov-state assumption in Eq. (1) is falsified.","tokens_in":14403,"feed_emoji":"🤖","tokens_out":4841,"duration_ms":45055,"temperature":0.7,"pith_summary":"The paper is trying to show that a robot can learn the waiter task—carrying an object on a tray without it sliding—from very few demonstrations and then handle objects it never saw. The proposed method trains a team of value functions from end-effector trajectories labeled only with sparse slip/no-slip outcomes, then uses them inside a model-predictive controller that treats disagreement among the team as a risk signal. If this works, it matters because current analytic controllers need accurate object friction and inertia, while behavior cloning needs many demonstrations; this recipe promises fast teaching with minimal labeling. On a real Franka Panda arm, the paper reports success on novel convex and household objects after 50–100 demonstrations, and shows it can learn from a demonstrator with wrong friction assumptions.","feed_headline":"A robot waiter learns unseen objects from 50 demonstrations","feed_subtitle":"Offline value ensembles plus risk-aware control keep objects stable even when shape and material are new.","key_machinery":"The load-bearing object is an ensemble of value functions $V_{\\phi_i}$ trained with the Bellman error objective of Eq. (1) on transitions $(x_t, c, x_{t+1})$, where $x$ is the end-effector pose, twist, and acceleration and $c \\in \\{0, c_{\\text{friction}}\\}$ is a sparse slip label. Online, MPC rolls out trajectories and scores them with a pessimistic return $\\hat{G}_{\\text{pess}} = \\log\\sum_i \\exp(\\hat{G}_i / \\lambda)$ of Eq. (2), which downweights trajectories where ensemble members disagree; applying this pessimism only at the initial state rather than pointwise is what avoids the over-conservatism identified in prior offline RL work. The value functions are not goal-conditioned: goal reaching comes from an $L^2$ cost, while the learned term supplies long-horizon slip avoidance.","core_discovery":"The paper's central claim is that the robot waiter task can be solved by offline value-function learning combined with online model-predictive control, without ever giving the learner the object's physical properties or dense costs. An ensemble of value functions is trained on end-effector states and sparse slip/no-slip labels, and the controller's trajectory-return estimate is made conservative through a log-sum-exp aggregation over the ensemble. This conservative value MPC (CV-MPC) is reported to match a friction-aware MPC demonstrator with exact object knowledge, beat the same demonstrator when its friction estimate is biased, and generalize to eight convex test objects and five household objects on a real Franka Panda arm, with success rates of about 73–95% and 80% or above respectively.","pith_inferences":["If the core assumption holds, the same recipe should extend to other contact-rich dynamic skills—such as carrying open containers, sweeping, or pushing—where end-effector kinematics plus sparse failure labels carry enough information; the paper only demonstrates the waiter task.","Human teleoperation would likely work as a demonstration source, since the learning pipeline never needs expert actions or dense labels; the paper tests only with an algorithmic demonstrator.","A direct stress test of the Markov assumption would hold end-effector trajectories fixed while varying object surface roughness or internal mass distribution; if slip prediction degrades sharply, the value-function state must be augmented with object properties.","The log-sum-exp pessimism over the ensemble is one possible uncertainty aggregator; alternatives such as max or quantile aggregation could be compared without changing the rest of the framework."],"forward_implications":["A practitioner with only end-effector demonstrations and sparse failure labels can deploy dynamic manipulation on hardware at 50 Hz using a single GPU, without estimating object friction or inertia.","The method can improve over a suboptimal demonstrator: when the demonstrator uses an incorrect friction coefficient, the learned value functions still recover higher success rates.","Learned value functions transfer to unseen objects: real-world tests report 73–95% success on eight convex shapes after training on a cube, and 80% or above on household objects after training on five hard convex objects.","The end-effector rotation is the critical observation; rotation alone outperforms full position-plus-rotation when training and test start positions differ, suggesting the learned skill is about tilt and motion, not absolute location.","Because the approach integrates as a cost term in off-the-shelf MPC, it can inherit constraint satisfaction, smooth joint-space motion, and the optimizer's safety properties."],"supporting_citations":[{"why":"Supplies the GPU-accelerated sampling-based MPC framework used for online deployment and the baseline trajectory-return cost.","marker":"[6]"},{"why":"Provides the friction-cone MPC formulation that generates the algorithmic demonstrations used for training.","marker":"[16]"},{"why":"Supplies the initial-state pessimism idea that the paper uses to avoid over-conservative value predictions.","marker":"[24]"},{"why":"Motivates blending value estimates at all steps in the MPC horizon to mitigate model errors.","marker":"[10]"},{"why":"Provides the conservative offline RL perspective that inspires penalizing out-of-distribution states.","marker":"[12]"},{"why":"Supports the use of an ensemble of value functions to capture epistemic uncertainty from limited data.","marker":"[22]"},{"why":"Defines the robot waiter problem and the analytical contact/friction modeling that CV-MPC aims to replace.","marker":"[5]"},{"why":"Frames the offline RL problem and the challenge of limited data coverage that motivates the pessimistic MPC objective.","marker":"[11]"}],"fun_headline_variants":["Robot waiter carries novel objects from just 50 demos","Offline value ensembles let robot waiter handle unseen shapes","Conservative MPC helps robot waiter generalize to new objects","Few-shot robot waiter: RL values guide risk-aware transport","50 demonstrations teach robot waiter to move unknown items"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned value functions see only the end-effector's motion and a sparse slip/no-slip label, so the whole method assumes these observations are enough to predict when an object will slip; if slip depends on object-specific properties such as contact patch, surface roughness, or internal mass distribution that the end-effector kinematics do not capture, transfer to novel objects will fail.","fun_headline_variants_meta":{"raw":{"variants":["Robot waiter carries novel objects from just 50 demos","Offline value ensembles let robot waiter handle unseen shapes","Conservative MPC helps robot waiter generalize to new objects","Few-shot robot waiter: RL values guide risk-aware transport","50 demonstrations teach robot waiter to move unknown items"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1712,"prompt_tokens":907,"completion_tokens":805,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":729}},"tokens_in":523,"tokens_out":805,"duration_ms":7466,"temperature":1.0,"reasoning_tokens":729,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:36:09.009558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a value-function ensemble on a cube with a fixed surface friction, then test on objects that produce identical end-effector trajectories and nominal friction but differ in internal mass distribution or contact-patch roughness; if success rates collapse while end-effector kinematics are held fixed, the Markov-state assumption in Eq. (1) is falsified.","supporting_citations":[{"cited_title":"Keep it upright: Model predictive control for nonprehensile object transportation with obstacle avoidance on a mobile manipulator,","cited_arxiv_id":null,"evidence_quote":"Provides the friction-cone MPC formulation that generates the algorithmic demonstrations used for training."},{"cited_title":"Bellman- consistent pessimism for offline reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the initial-state pessimism idea that the paper uses to avoid over-conservative value predictions."},{"cited_title":"Uncertainty-based offline reinforcement learning with diversified q-ensemble,","cited_arxiv_id":null,"evidence_quote":"Supports the use of an ensemble of value functions to capture epistemic uncertainty from limited data."},{"cited_title":"A shared-control teleoperation architecture for nonprehensile object transportation,","cited_arxiv_id":null,"evidence_quote":"Defines the robot waiter problem and the analytical contact/friction modeling that CV-MPC aims to replace."}],"review_version":1}