{"id":"aa002ab9-a6a7-4aca-9797-7e1ecdd2e8f1","arxiv_id":"1908.07088","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Contextual bandit algorithms (epsilon-greedy and LinUCB) with SPANet features let a feeding robot acquire previously unseen foods after about ten attempts per food.","lead":"Robot-assisted feeding systems can learn the best skewering strategy for a new food in roughly ten failed attempts, using standard bandit algorithms on top of a pretrained visual model. This is a step toward assistive robots that adapt to each user's foods without expensive retraining.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed convergence 'within 10 attempts' for the pretrained-θ scenario is not actually shown: Experiment 2's cumulative-loss plots (Fig. 7) begin at attempt 20, and the first 20 adaptation attempts are omitted from quantitative analysis.","rationale":"The reader's linearizability concern (§III-B) is real and is made more urgent by Figure 2, which shows SPANet—a linear head on the frozen penultimate features—performing worse than random for excluded banana. However, the empirical convergence claim can hold even under misspecification, particularly because Algorithm 1's resampling-on-success rule repeats the same context on failure, allowing the learner to memorize which action works for that specific food. In that regime the linear map is not the critical load-bearing element for the stated speed-of-convergence claim. What is critical is that the exact scenario in the abstract—pretrained θ plus a novel food—is actually measured over the first attempts. Figure 7's x-axis starts at attempt 20, so the claimed 10-attempt convergence in that scenario is not shown. Thus the more load-bearing issue is evidential: the strongest form of the claim is unsupported by the reported plots, even though a fuller log might rescue it. Keeping the verdict conditional on full release and re-analysis of those early attempts is the defensible position; nothing in the current paper justifies either ACCEPT or REJECT beyond CONDITIONAL.","tokens_in":12232,"tokens_out":9973,"duration_ms":106996,"concrete_test":"Obtain or release the full per-attempt logs for Experiment 2 (banana and carrot, all three algorithms) and plot cumulative loss and the identity of each selected action for attempts 1–30, together with a random-action baseline. Record the first attempt at which the selected action is and remains in the best-strategy set. If that attempt is at most 10 for the pretrained-θ LinUCB runs, the claim is supported; if substantial suboptimal actions occur during attempts 1–20, the 'within 10 attempts' claim should be withdrawn or restricted to Experiment 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's strongest quantitative claim—that with a model trained on thousands of dissimilar items, ε-greedy and LinUCB converge to the best strategy for previously unseen food within about 10 attempts—is not demonstrated by the plotted real-robot data. The only experiment run with a pretrained θ is Experiment 2 (§IV-B(d)), but Figure 7 plots cumulative loss only for attempts 20–30, and per the caption attempts 20–25 are on a different, previously seen food item. The first 20 attempts on the novel food, where 'within 10 attempts' convergence would have to appear, are entirely absent from the figure and from the quantitative discussion. Figure 8 shows UCB estimates at attempts 1, 10, 20, and 30 for a single illustrative banana run, but provides no policy-selection or cumulative-loss summary. Moreover, Algorithm 1 re-samples the context only after success, so failures repeat the same context; with only six actions, repeated probing of one food item can produce apparent per-item convergence without any contextual generalization. Section V's statement that 'both experiments suggest' the 10-attempt convergence therefore overreaches: Experiment 1 initializes θ=0, and Experiment 2's plotted evidence starts after the adaptation window has closed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online learning framework for robot-assisted feeding in which bite acquisition is cast as a contextual bandit problem. Food images are featurized by the penultimate layer of SPANet, and the expected loss of each of six manipulation strategies is modeled as a linear function of those features. The authors use an importance-weighted linear regression oracle, test epsilon-greedy and LinUCB exploration strategies, tune hyperparameters in a simulated environment built from prior data with a doubly robust loss estimator, and then evaluate on a real robot. Experiment 1 starts from a zero initial theta on apple, banana, and grape; Experiment 2 starts from a pretrained theta and tests banana and carrot. The central claim is that, even for previously unseen food items requiring very different strategies, the algorithms converge to the best strategy within about 10 attempts.","tokens_in":12475,"tokens_out":4059,"duration_ms":45274,"significance":"If validated, the framework would be a useful step toward data-efficient adaptation in assistive feeding, reducing the need to pre-collect food-specific training data. The paper has genuine strengths: the contextual-bandit formulation is clean, the use of a doubly robust estimator to build a simulation environment is appropriate, the algorithms are standard and clearly described, and the real-robot evaluation addresses a practically important problem. The main limitation is that the strongest quantitative claim, convergence within about 10 attempts for previously unseen foods, is not directly supported by the plotted real-robot data, and the experimental evidence is thin in several load-bearing places. The paper is therefore plausible but needs substantial strengthening of the empirical support before the central claim can be accepted.","major_comments":[{"comment":"The claim that both experiments show convergence 'within 10 attempts' is not supported by the plotted Experiment 2 data. Figure 7 shows cumulative loss only for attempts 20 through 30, and the caption states that attempts 20 through 25 were performed on a previously seen food item. The first 20 attempts on the novel food, which is exactly the window in which the 'within 10 attempts' convergence would need to appear, are omitted from both the figure and the quantitative discussion. Section V's statement that 'both experiments suggest that this convergence could happen within 10 attempts' therefore overreaches; at most the data show stable behavior after attempt 25 on banana and after some unreported point on carrot.","section":"Section V / Figure 7"},{"comment":"The hyperparameters (epsilon, alpha, lambda, and d) are tuned in simulation using banana, apple, and grape as the excluded foods, and the same three foods are then used in the real-robot Experiment 1. This means those foods are not genuinely 'previously unseen' at the system level: information about their success-rate structure has influenced the choice of exploration parameters and regularization. The abstract's claim about adapting to previously unseen foods is weakened because the tuner had access to the test foods. A cleaner protocol would tune on a separate held-out set of foods and reserve the three test foods entirely, or would report sensitivity of the real-robot results to the chosen hyperparameters.","section":"Section IV-A / Figure 5"},{"comment":"The linear realizability assumption, E[ct | at, xt] = theta_a^T phi(xt), is load-bearing: if the true expected loss is not linear in the SPANet features for an out-of-distribution food, no amount of online data can make the weighted least-squares oracle converge to the correct policy. The paper justifies this only by citing SPANet's success, but the entire point of the paper is to handle foods whose optimal strategies differ from the training distribution. The authors should provide a direct check, for example residual or calibration plots on held-out food data, or a comparison of the linear model against a nonparametric estimate for the foods used in the experiments. Without such evidence, the convergence claims rest on an untested functional form.","section":"Section III-B / Equation (2)"},{"comment":"The experimental section does not describe repeated trials for either experiment, despite Figure 7 displaying a '3-Trial Range' that is never explained in the text. If the range is over multiple runs, the number of runs and the protocol should be stated. If it is a within-run quantity such as a moving window, that should be clarified. As written, the reader cannot determine whether the reported cumulative-loss curves are single trajectories or averages, and the paper's comparative statements about greedy, epsilon-greedy, and LinUCB are therefore not backed by statistical evidence.","section":"Section IV-B / Figures 6 and 7"},{"comment":"Because the context is resampled only after a success, failures repeat the same context. In Experiment 2, where each trial uses a single food item and failures dominate early attempts, the algorithm may appear to converge for that specific food item without demonstrating any contextual generalization to unseen foods. This is worth acknowledging explicitly; as presented, the evidence supports per-item adaptation more strongly than it supports generalization across novel food categories.","section":"Algorithm 1 / Section III-A"}],"minor_comments":[{"comment":"Equation (2) is written as a squared-loss objective but is not a complete definition of the weighted least-squares estimate; the notation should make clear that theta_hat minimizes this expression over all actions and that the sum runs over collected samples, not over the feature dimension.","section":"Section III-B / Equation (2)"},{"comment":"The vertical line in Figure 6 is described only in the caption as the point after which 100% of strategies selected were among the best strategies. The text should define what 'best strategy set' means formally and explain how the line is computed, since the convergence claim in Section V depends on this definition.","section":"Figure 6"},{"comment":"The x-axis labels in Figure 7 read 'Banana Acquisition Attempt' and 'Carrot Acquisition Attempt', but the plotted range is 20-30 and includes five attempts on a different previously seen food. The axis and caption should be clarified to distinguish the test-food attempt index from the total attempt index.","section":"Figure 7"},{"comment":"The procedure states that 'we removed and replaced the food item only after a successful acquisition,' but Experiment 2's protocol of 20 attempts followed by 5 attempts on a previously seen food and then 5 more attempts is not described as repeated. Please clarify whether the food item is replaced after every attempt or only after success, and how the fixed 20-attempt count is maintained.","section":"Section IV-B(b)"},{"comment":"LinUCB is cited in different forms across the paper: [10] is used for the UCB formula, [45] for the original LinUCB paper, and [50] for the regret bound. Unifying these citations would help the reader trace the exact algorithm and guarantees being used.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central idea is sound and likely of interest to the assistive-robotics community, but the strongest quantitative claim is not backed by the reported data: the first 20 adaptation attempts in Experiment 2 are missing from the cumulative-loss plots, and the hyperparameter tuning uses the same foods that later appear as the 'unseen' test foods. These are fixable with additional experiments or a revised, more limited claim, so I recommend major revision rather than rejection. I would also encourage the editor to ask the authors to make the raw per-attempt data available, since the current figures do not permit independent assessment of variance or convergence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent framework contribution with a weak empirical section, and the claim that a pretrained model can adapt to unseen foods within 10 attempts is not supported by the evidence presented.\n\nThe new idea is treating bite acquisition as a linear contextual bandit on SPANet penultimate-layer features, plus a doubly robust simulator for tuning hyperparameters. That's clean and reusable. The real-robot experiments do show that LinUCB can adapt to banana and carrot, whose optimal strategies differ from the training distribution, while not completely forgetting previously seen foods. The experiments are suggestive and align with the prior SPANet results.\n\nNow the soft spots. The most serious, as you may have seen from the stress-test, is that Experiment 2's cumulative-loss plots begin at attempt 20. The first 20 attempts on the novel food—where 'within 10 attempts' convergence would have to appear—are never shown or summarized. The five attempts on a previously seen food in between further muddy the picture. Figure 8 shows a single illustrative banana run, but that's not a loss curve. So the abstract's 'quickly converge' is an overreach relative to what's plotted.\n\nThere's also tuning leakage: the hyperparameters are tuned in simulation on banana, apple, and grape, and those same foods are used in Experiment 1. The effect is probably modest—only a few scalars—but it should be disclosed. More importantly, the experiments appear to be single runs; Figure 7's '3-Trial Range' is never explained in the text. No code or data are released, which makes the simulation-to-real story hard to verify.\n\nThe linear realizability assumption is a real limitation, but the paper states it plainly, and the empirical results suggest it holds for the tested foods. Not fatal.\n\nNet: this is a solid framework paper that deserves a serious referee, but it needs a revision with the full attempt curves, repeated trials, and a more careful statement of what is demonstrated. The '10 attempts' claim should be scaled back or supported with data. This is for robotics researchers working on adaptive manipulation and assistive feeding, and to a lesser extent contextual bandit practitioners. I'd send it to peer review with a request for major revision and artifact release.","headline":"A clean bandit framework for adaptive feeding, but the paper's headline convergence claim is not actually backed by the plotted data.","tokens_in":13046,"tokens_out":4165,"would_cite":false,"duration_ms":41326,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-14T12:27:05.740032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}