{"id":"7dd02761-6373-46e0-88d3-53ffaf473cea","arxiv_id":"2505.10359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hierarchical language-conditioned policy that synthesizes an adaptively selected novel view reaches 90.4% single-task success and 2.93 average consecutive-task length on CALVIN, ahead of prior non-foundation-model methods.","lead":"This paper introduces a robot policy that generates a second, synthetic camera view of the scene from an automatically chosen angle and uses it, together with the original views and a language instruction, to choose manipulation actions. The system reports the best published results among non-foundation-model methods on the CALVIN benchmark and shows larger gains in a small real-robot test.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adaptive-angle advantage is not statistically established: fixed θ=20° nearly matches Eq. (1) (2.89 ± 0.06 vs 2.93 ± 0.04), so the central novelty—adaptive viewpoint selection—remains unsupported without sensitivity analysis.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing point: Eq. (1) is an ad hoc linear heuristic, and Table II shows a fixed θ=20° nearly matches the adaptive result. My stress-test agrees and sharpens the concern. The central claim of the paper, as stated in the abstract and Section IV, is that adaptive novel-view synthesis is what yields 90.4% success and 2.93 average horizon length, and that this paradigm—not just novel-view synthesis per se—is the contribution. The evidence for the adaptive component rests entirely on Table II, where the margin over the best fixed angle is smaller than the reported error bars. Moreover, the w/o NS ablation (Table III) shows that removing synthesis entirely drops average length to 2.71, so the gap between no synthesis and fixed 20° (0.18) is more than four times the gap between fixed 20° and adaptive (0.04). This suggests the real effect being measured is synthesis itself, not the adaptive selection rule. The unspecified 'small location perturbation' and the absence of any sensitivity analysis over w1 and w2 reinforce the concern that the specific coefficients may be tuned to CALVIN's depth statistics. The Table III inconsistency between w/o NS and w/o FD rows is a separate but related reporting problem that undermines confidence in the ablation story. None of this makes the paper's empirical claims fraudulent or even necessarily wrong—the comparisons against external baselines on a public benchmark are appropriate—but it does mean the headline contribution is not yet supported at the usual standard of evidence. The correct verdict remains CONDITIONAL: the authors should supply a proper significance test and sensitivity analysis, plus clarify the perturbation and the Table III rows. I therefore recommend no change to the reader's verdict, while noting that if the proposed test shows no significant adaptive advantage, the paper's contribution should be substantially reframed.","tokens_in":12425,"tokens_out":3387,"duration_ms":35539,"concrete_test":"Re-run the CALVIN evaluation with at least 10 seeds for three configurations: (i) Eq. (1) with w1=14, w2=39; (ii) fixed θ=20°; (iii) fixed θ=15° and θ=25°. Use identical GenWarp, cycle-consistent VAE, and policy training. Compute paired two-sided bootstrap or permutation confidence intervals on the difference in average horizon length and on Horizon 4/5 success rates. If the 95% CI for (i) minus (ii) includes zero or is negative, the claim that adaptive viewpoint selection outperforms a fixed moderate angle is unsupported, and the contribution should be reframed as 'synthesizing any informative novel view helps' rather than 'adaptive selection is essential.' Also report the distribution of dcam-ori over CALVIN episodes to check whether the linear rule is well-matched to a narrow depth range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adaptively selected novel views drive NVSPolicy's success. Table II is the only direct evidence, and it is weak: the adaptive rule Eq. (1) reaches 2.93 ± 0.04 average horizon length, while fixed θ=20° reaches 2.89 ± 0.06. The difference is 0.04, within reported error bars, and at Horizon 2 the fixed angle is actually higher (71.2 vs 70.4). No significance test, paired comparison, or seed-level variance is reported, so the superiority of Eq. (1) over a constant angle is not demonstrated. The rule itself is an unexplained linear heuristic: no derivation, no units for dcam-ori, no discussion of the depth range over which w1=14 and w2=39 were tuned, and the 'small location perturbation' is unspecified. If w1 and w2 were tuned to CALVIN's depth statistics, the generalization claim beyond CALVIN is at risk. Additionally, Table III reports nearly identical values for the w/o NS and w/o FD rows (e.g., Horizon 1: 88.3 vs 88.3; Horizon 5: 26.1 vs 26.1), likely a copy-error or uncontrolled ablation, which further weakens the decomposition of contributions. The paper does support that novel-view synthesis itself helps (w/o NS vs fixed angles), but the specific adaptive-selection contribution is the least-supported and most load-bearing component of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NVSPolicy, a language-conditioned robotic manipulation policy that augments the gripper-camera observation with a novel-view image synthesized by GenWarp at an adaptively selected viewpoint. A cycle-consistent VAE disentangles the visual features into a semantic feature and a remaining feature; the semantic feature drives a high-level meta-skill selector, while the remaining feature drives a low-level action estimator. The method also includes keyframe selection and policy distillation for efficiency. The paper reports state-of-the-art results on CALVIN (Horizon-1 success 90.4%, average horizon length 2.93), ablation studies isolating the contribution of novel-view synthesis, feature disentanglement, and policy distillation, and a small real-robot comparison against SPIL.","tokens_in":12889,"tokens_out":5552,"duration_ms":48334,"significance":"If the reported gains are statistically robust, the paper makes a useful contribution: it demonstrates that a pretrained single-image novel-view synthesizer can provide task-relevant context for language-conditioned policy learning, and that disentangling semantic from remaining features is a sensible way to mitigate synthesis artifacts. Evaluation on the public CALVIN benchmark against several recent non-foundation-model baselines is a strength, as is the inclusion of ablation and real-robot studies. However, the statistical support for the key adaptive-viewpoint mechanism is currently weak, and an internal inconsistency in Table III limits what can be concluded about the feature-disentanglement component.","major_comments":[{"comment":"The rows 'w/o NS' and 'w/o FD' report identical success rates at Horizons 1, 2, 3, and 5 (88.3/88.3, 68.7/68.7, 51.3/51.3, 26.1/26.1) but different average horizon lengths (2.71 vs. 2.73), and the text states a reduction to 2.74 for w/o FD. This internal inconsistency makes it impossible to evaluate the contribution of feature disentanglement, which is a load-bearing component of the proposed method. Please correct the table and text, and either rerun the w/o FD ablation or report the actual numbers with the correct averages.","section":"Table III / Sec. IV-D"},{"comment":"The adaptive viewpoint rule is the central novelty of the paper, but Table II shows that the fixed angle θ=20° achieves an average horizon length of 2.89 ± 0.06 versus 2.93 ± 0.04 for the adaptive rule, and at Horizon 2 the fixed angle is higher (71.2 vs. 70.4). No significance test, paired comparison, or per-seed breakdown is reported, so the evidence that adaptive selection outperforms a fixed viewpoint is not demonstrated. Please provide statistical tests (e.g., paired tests across seeds or task sequences) and report per-seed results, or temper the claim accordingly.","section":"Table II / Sec. IV-D"},{"comment":"The linear heuristic θ = −w1·d_cam-ori + w2 with w1=14 and w2=39 is introduced without derivation, without specifying the units or valid range of d_cam-ori, and without sensitivity analysis. The 'small location perturbation' added to the sampled viewpoint is also unspecified. If the coefficients were tuned on CALVIN's depth statistics, the generalization claim beyond CALVIN is not supported. Please justify the heuristic, report the depth range over which it was designed or tuned, and include a sensitivity study over w1, w2, and the perturbation magnitude.","section":"Sec. III-B, Eq. (1)"}],"minor_comments":[{"comment":"The real-robot evaluation uses only 10 trials per task and reports a single baseline (SPIL) with no variance, error bars, or statistical test; please state this limitation explicitly and, if possible, report per-trial outcomes or confidence intervals.","section":"Sec. IV-E / Table IV"},{"comment":"There are several typographical and grammatical errors, e.g., 'can generation effort' in Sec. I, 'a adaptive novel viewpoint' in Sec. I, 'mechansim' in Sec. III-C, and 'we does not consider' in Sec. III-D; a careful proofread is needed.","section":"Sec. I, Sec. III-C, Sec. III-D"},{"comment":"Several symbols in the notation list (MI(·), LKD, m, n, τ, K, p(It−1, It)) are not used in the main text, which appears to be leftover material from an earlier draft; please remove them or introduce them in the appropriate sections.","section":"Appendix VI-B"},{"comment":"The arrows and labels in Figs. 2 and 4 are dense and partly illegible, especially the feature-exchange paths in the forward and reverse processes of the cycle-consistent VAE; please enlarge fonts or add a clearer explanatory caption.","section":"Figs. 2 and 4"}],"recommendation":"major_revision","confidential_remarks":"The main concern is that the paper's central novelty, adaptive viewpoint selection, is currently supported only by a small, non-significant-looking difference over fixed θ=20° in Table II, and the ablations in Table III contain an apparent copy-paste error between the w/o NS and w/o FD rows. Both should be addressed with re-runs and proper statistical reporting before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read NVSPolicy. The useful take: this is a legitimate integration paper—GenWarp-style NVS, cycle-consistent VAE disentanglement, and a hierarchical policy—evaluated on CALVIN with a real-robot sanity check. If the headline numbers are reproducible, it is a reasonable incremental advance, not a breakthrough. The paper earns credit for evaluating against external baselines on a public benchmark and for running ablations on each component. The circularity risk is low; nothing is fitted to the success metric except hyperparameters.\n\nThe soft spots are exactly where the reader and stress test point. The adaptive-viewpoint rule Eq. (1) is a two-constant linear heuristic with no derivation, no units, and no sensitivity analysis. Table II shows the fixed angle θ=20° at 2.89 ± 0.06 versus 2.93 ± 0.04 for the adaptive rule; the difference is within error bars and at Horizon 2 fixed θ20 is higher. The paper nevertheless claims the adaptive strategy consistently outperforms. That overstates the evidence. The 'small location perturbation' is unspecified. If w1 and w2 were tuned to CALVIN depth statistics, the generalization claim weakens.\n\nTable III has a concrete reporting problem: w/o NS and w/o FD rows are nearly identical (Horizon 1–3 and 5 exactly the same), and the text says w/o FD drops average length to 2.74 while the table says 2.73. That reads like a copy-paste error or an uncontrolled ablation; either way it needs fixing. The real-robot experiment uses 10 trials per task, which is thin, though the gap over SPIL is large enough to be suggestive.\n\nMy verdict: conditional. The central empirical claim—that synthesized novel views help—is plausible and supported. The stronger claim that adaptive selection is the source of the advantage is not established with the reported statistics. The paper deserves a serious referee: request code, seed-level variance, significance testing for Table II, corrected Table III, a sensitivity analysis for Eq. (1), and a clearer description of the viewpoint perturbation. I would bring it to reading group as an example of where NVS meets manipulation, but cite it mainly for the integration and the benchmark numbers, not for the adaptive rule.","headline":"A plausible integration paper whose adaptive-viewpoint novelty is not statistically supported; deserves revision, not rejection.","tokens_in":13292,"tokens_out":1944,"would_cite":true,"duration_ms":17539,"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":"Adaptive novel-view synthesis lifts language-conditioned robot success to 90.4%.","keywords":["novel-view synthesis","language-conditioned policy learning","hierarchical policy","cycle-consistent VAE","feature disentanglement","CALVIN benchmark","imitation learning","robot manipulation"],"falsifier":"Move the wrist camera to a different mounting height on the same simulator and rerun the identical pipeline with $w_1=14$, $w_2=39$: if the average horizon length is not above the best fixed angle, roughly 20 degrees, or if a grid search over the coefficients shows a broad plateau, then the adaptive viewpoint rule is not doing the causal work claimed.","tokens_in":12213,"feed_emoji":"🤖","tokens_out":7081,"duration_ms":65660,"temperature":0.7,"pith_summary":"NVSPolicy claims that a robot following natural-language instructions can be made more reliable by first generating an extra camera view of the scene from a viewpoint chosen on the fly, then feeding that synthetic view through a two-part encoder. The semantic part of the encoded features chooses which high-level skill to execute, while the remaining part predicts low-level motion, so flaws in the generated image do not corrupt the skill choice. On the CALVIN long-horizon benchmark the method reaches an average success rate of 90.4% and completes 2.93 consecutive tasks on average, above the best prior methods that do not use large foundation models, and it also beats a baseline on a physical robot. The point of the work is that imperfect generative images are usable for control if the policy separates what is in the scene from everything else.","feed_headline":"Robot policy gains 90% success by synthesizing a second camera view","feed_subtitle":"The depth-chosen extra view and split features push CALVIN horizon length to 2.93 and real-robot success to 44%.","key_machinery":"The load-bearing object is the adaptive viewpoint rule $\\theta = -w_1 d + w_2$ with $w_1=14$ and $w_2=39$, where $d$ is the average scene depth; this sets how far the synthesized camera swings away from the observed view. GenWarp, a pre-trained single-image generative warping model, then renders that novel view. Around this sit a cycle-consistent VAE that disentangles each image into semantic and remaining latent features, a hierarchical policy in which the semantic features drive meta-skill selection and the remaining features drive the action estimator, and two efficiency mechanisms, keyframe-only synthesis and policy distillation, that let a student network estimate semantic features without running the generative model at every step.","core_discovery":"The paper's central claim is that synthesizing a novel viewpoint from the wrist camera, with the viewpoint chosen by a simple depth-based rule, gives a language-conditioned policy enough extra visual context to generalize better across long task chains. To make the imperfect synthetic image safe, the authors train a cycle-consistent VAE that splits each image into a semantic feature, shared across views and used for meta-skill selection, and a remaining feature, used for action estimation. Ablations support the claim: removing novel-view synthesis lowers the average horizon length from 2.93 to 2.71, removing feature disentanglement lowers it to 2.73, and replacing the adaptive angle with the best fixed angle of 20 degrees lowers it to 2.89. The authors conclude that adaptive novel-view synthesis plus disentangled features is the source of the improvement.","pith_inferences":["A natural next test is whether the viewpoint rule $\\theta = -14 d + 39$ transfers: on a benchmark with a different camera rig or depth distribution the same coefficients may need re-tuning, and a grid search would show whether the adaptive advantage is a sharp optimum or a broad plateau.","Because the policy deliberately ignores the synthesized image's remaining feature, the same semantic-versus-remaining disentanglement could admit other imperfect generative inputs, such as inpainted occlusions or hallucinated goal states, without retraining the action estimator.","Keyframe selection plus distillation implies the synthesized view acts mainly as an auxiliary training signal, which suggests a fully self-supervised variant could train the student directly on sampled viewpoints and remove the teacher at deployment."],"forward_implications":["If the central claim holds, a robot policy can safely consume generative images: semantic features are robust to synthetic artifacts while action-relevant details come from real cameras.","Removing the synthesized view costs about 0.22 tasks of average horizon length, so the novel-view context is doing real work for long task chains.","Adaptive viewpoint selection outperforms every fixed angle tested, with the largest margin at horizon 5, 32.8% versus 29.2% for 20 degrees.","Policy distillation can replace explicit synthesis during inference, meaning the generative model acts as a training-time teacher rather than a runtime cost.","The same architecture transfers to a physical robot, with average success 44% versus 8% for the compared baseline across five tasks."],"supporting_citations":[{"why":"Supplies the pre-trained generative warping model, GenWarp, that renders novel views from a single input image.","marker":"[8]"},{"why":"Supplies the cycle-consistent VAE training objective used to disentangle semantic from remaining visual features.","marker":"[35]"},{"why":"Provides the CALVIN benchmark, task sequences, and evaluation protocol for long-horizon, language-conditioned manipulation.","marker":"[36]"},{"why":"Is the closest hierarchical baseline and supplies the rule-based meta-skill labeling and selector design that NVSPolicy adapts.","marker":"[6]"},{"why":"Supplies the hierarchical action-estimator architecture that the low-level policy builds on, and serves as a baseline.","marker":"[7]"},{"why":"Is the strongest diffusion-based baseline used for leaderboard comparison.","marker":"[20]"},{"why":"Provides the pre-trained MiniLM language encoder that converts instructions into the language feature.","marker":"[37]"}],"fun_headline_variants":["Adaptive novel-view synthesis boosts robot policy to 90% CALVIN success","Depth-chosen synthetic view lifts robot policy success to 90.4% on CALVIN","Synthetic second view with split features lifts CALVIN success to 90.4%","Adaptive camera view synthesis helps robot policy hit 90.4% on CALVIN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-set linear viewpoint rule $\\theta = -w_1 d + w_2$, with $w_1=14$ and $w_2=39$, generalizes beyond CALVIN's camera setup; it is an ad hoc heuristic with no derivation, and a fixed 20-degree viewpoint nearly matches its average horizon length, 2.89 versus 2.93, so the adaptive advantage rests on coefficients that could be benchmark-specific.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive novel-view synthesis boosts robot policy to 90% CALVIN success","Depth-chosen synthetic view lifts robot policy success to 90.4% on CALVIN","Synthetic second view with split features lifts CALVIN success to 90.4%","Adaptive camera view synthesis helps robot policy hit 90.4% on CALVIN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3463,"prompt_tokens":999,"completion_tokens":2464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":2370}},"tokens_in":615,"tokens_out":2464,"duration_ms":15083,"temperature":1.0,"reasoning_tokens":2370,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:09:44.756707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Move the wrist camera to a different mounting height on the same simulator and rerun the identical pipeline with $w_1=14$, $w_2=39$: if the average horizon length is not above the best fixed angle, roughly 20 degrees, or if a grid search over the coefficients shows a broad plateau, then the adaptive viewpoint rule is not doing the causal work claimed.","supporting_citations":[{"cited_title":"Genwarp: Single image to novel views with semantic- preserving generative warping","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained generative warping model, GenWarp, that renders novel views from a single input image."},{"cited_title":"Disentangling factors of variation with cycle-consistent variational auto-encoders","cited_arxiv_id":null,"evidence_quote":"Supplies the cycle-consistent VAE training objective used to disentangle semantic from remaining visual features."},{"cited_title":"Language-conditioned imitation learning with base skill priors under unstructured data","cited_arxiv_id":null,"evidence_quote":"Is the closest hierarchical baseline and supplies the rule-based meta-skill labeling and selector design that NVSPolicy adapts."},{"cited_title":"Language control diffusion: Efficiently scaling through space, time, and tasks","cited_arxiv_id":null,"evidence_quote":"Is the strongest diffusion-based baseline used for leaderboard comparison."},{"cited_title":"Minilm: Deep self-attention distillation for task-agnostic com- pression of pre-trained transformers","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained MiniLM language encoder that converts instructions into the language feature."}],"review_version":1}