{"id":"531f7f5f-9dbf-44a8-acb2-6202ee4c93f2","arxiv_id":"2505.04619","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"By summing multi-view features and adding single-view features as actor-critic augmentations, MAD produces manipulation policies that learn faster and tolerate missing cameras in simulation.","lead":"Robots trained on multiple camera views usually break down if one camera fails at deployment. This paper introduces a training method that merges views during learning while also teaching the robot to keep working from any single view, and reports gains on 20 simulated manipulation tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MAD's reported gains hinge on alpha=0.8, which is tuned on the same tasks used for the headline Meta-World benchmark; without a disclosed held-out tuning protocol, the 30%/36% margins are not independently verified.","rationale":"The paper is a solid empirical RL contribution: MAD merges multi-view features through summation and applies SADA-style selective augmentation with single-view features, supported by 20 tasks, 5 seeds, 95% confidence intervals, component ablations, merge-method comparisons, and occlusion/modality tests. The central claim is that MAD improves sample efficiency while remaining robust to camera-view reduction. After reading carefully, the most load-bearing concern is not the mere transfer of the SADA recipe but the undisclosed and benchmark-sensitive tuning of alpha. The paper's own alpha ablation (Figure 5) shows a sharp cliff: alpha=0.5 and alpha=1.0 both give roughly 26% average success while alpha=0.8 gives 56%. That is a 30-point swing from a single hyperparameter. The authors say alpha is tuned but give no protocol, and the tasks used for tuning are the same 5 hard tasks that appear in the main Meta-World aggregate. This creates a selection-on-test risk: the reported 30% improvement on Meta-World may partly reflect picking alpha that works on those exact tasks. ManiSkill3 is a genuinely separate benchmark and alpha=0.8 transfers there, which is evidence of some robustness, but the sensitivity remains high enough that without a held-out tuning procedure the headline margins are conditional. The reader's weakest_assumption identified the SADA-recipe transfer as the key premise. I partially agree: the existence of the 'MAD - Naive Both' ablation does provide evidence that the selective-target formulation matters, but that ablation is confounded because it changes both the target rule and the weighting. My concern is more specific: the claimed stabilization could come from alpha weighting rather than from SADA's target-selection principle, and the alpha choice is not independently validated. This does not change the reader's CONDITIONAL verdict; it reinforces it. The paper should disclose the alpha-tuning protocol and ideally demonstrate that alpha=0.8 is stable across held-out task sets before the numerical margins are taken at face value. The method has merit and the empirical work is substantial; the missing piece is a rigorous hyperparameter-selection assessment.","tokens_in":19073,"tokens_out":8252,"duration_ms":83699,"concrete_test":"Run the alpha sweep {0.5, 0.7, 0.8, 0.9, 0.95, 1.0} on a held-out set of 5 Meta-World tasks disjoint from the 5 hard tasks used in Figure 5, fixing all other hyperparameters; then evaluate the alpha=0.8 model on the original 15-task Meta-World benchmark. If the optimal held-out alpha is not 0.8, or if the fixed-alpha model underperforms the re-tuned model by more than 5% average success, alpha selection is benchmark-dependent and the headline margins require per-benchmark tuning. In addition, run a 2x2 ablation crossing SADA target-selection with naive targets at alpha=0.8 and alpha=0.5 on the same 5 hard tasks to separate the contributions of SADA's target rule from alpha weighting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 defines the MAD objective as a convex combination of a merged-view (unaugmented) loss and single-view feature-augmented losses, controlled by a single hyperparameter alpha. The paper states that alpha is tuned (Section 4.3) and fixes alpha=0.8 in Table 1. Figure 5's alpha ablation shows this choice is load-bearing: alpha=0.5 (the original SADA weighting) and alpha=1.0 (merged-only) both collapse to around 26% average success on the 5 hard Meta-World tasks, while alpha=0.8 reaches 56%. The tuning protocol is undisclosed: no search range, number of seeds, or selection criterion is reported. Critically, the alpha ablation and the headline Meta-World results in Figure 4 both include the same 5 hard tasks (Assembly, Hand Insert, Pick Out of Hole, Pick Place, Push; Appendix D.2). If alpha was chosen by evaluating on those tasks, the reported Meta-World averages are partially selected on the evaluation set, making the sample-efficiency comparison optimistic. ManiSkill3 provides an out-of-benchmark transfer of alpha=0.8, which is suggestive, but the extreme sensitivity around alpha means a small change in weighting could erase the reported margins. In addition, the paper does not compare a naive multi-state objective with alpha weighting, so the stabilization attributed to the SADA recipe (Eqs. 1-2) is confounded with the alpha weighting; the 'MAD - Naive Both' ablation uses an unweighted naive objective and therefore cannot isolate the source of stability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAD (Merge And Disentangle), a modification of the DrQ visual RL algorithm for multi-camera robotic manipulation. Each camera view is encoded by a shared CNN and the resulting features are summed into a merged representation; the actor and critic are trained on the merged representation as an unaugmented stream and on each single-view feature as a feature-level augmentation, following the SADA recipe with an additional weighting hyperparameter alpha. The method is evaluated on 15 Meta-World and 5 ManiSkill3 tasks against MVD, VIB, MV-MWM, and Single Camera DrQ, with the central claim that MAD increases sample efficiency while remaining robust to reductions in input camera views. The paper also reports component ablations, alpha ablations, comparisons of merging methods, occlusion experiments, modality experiments, and scalability experiments with five views and mixed input resolutions.","tokens_in":19342,"tokens_out":4844,"duration_ms":46664,"significance":"If the reported results hold, MAD is a simple and lightweight contribution to multi-view visual RL: it needs no auxiliary losses, no additional learnable parameters, no ordered views, and no extra forward passes, while simultaneously improving sample efficiency and robustness to missing views. The empirical breadth is a genuine strength: 20 tasks, 5 seeds, 95% confidence intervals, component ablations, merge-method comparisons, occlusion and modality experiments, and an extended five-view experiment. The main result is also externally benchmarked against strong published baselines rather than only against self-derived variants. The central risks are the undisclosed alpha tuning protocol and the fact that the SADA stabilization recipe is imported from prior work without an ablation that isolates it from the alpha weighting; both are addressable within the manuscript's scope. The absence of real-robot validation is explicitly acknowledged in Section 6 and is not, by itself, a blocker for the stated simulation-based claims.","major_comments":[{"comment":"The choice alpha=0.8 is load-bearing, but the tuning protocol is not reported. Section 4.3 states only 'after tuning alpha', Table 1 fixes alpha=0.8, and Figure 5 (bottom) shows that alpha=0.5 and alpha=1.0 both average around 26% success on the 5 Meta-World hard tasks while alpha=0.8 reaches 56%. Since the alpha ablation in Appendix D.2 and the headline Meta-World results in Figure 4 both include the same 5 hard tasks (Assembly, Hand Insert, Pick Out of Hole, Pick Place, Push), selection of alpha on those tasks would make the Figure 4 averages partially selected on the evaluation set. Please disclose the search range, the number of seeds used for selection, the selection criterion, and ideally tune alpha on a separate task split or report sensitivity curves for all 15 tasks; the ManiSkill3 transfer of alpha=0.8 is suggestive but does not remove this concern.","section":"Section 4.3, Eq. (1)-(2), Table 1, Figure 5"},{"comment":"The ablation labeled 'MAD - Naive Both' does not isolate the SADA stabilization recipe from the alpha reweighting. MAD uses the SADA-structured losses with alpha=0.8, while 'MAD - Naive Both' is described only as training on merged and singular view features 'without using the MAD loss formulations'; no alpha-weighted version of the naive objective is reported. As a result, the comparison cannot distinguish whether the stability comes from the SADA target-selection structure (targets predicted from the unaugmented stream) or simply from the high weight on the unaugmented merged-view loss. Because the paper's central assumption is that the SADA recipe transfers from image-level augmentation to feature-level augmentation, please add an ablation that applies the same alpha weighting to a naive merged-plus-singular objective, or otherwise support the attribution of stability to the SADA recipe.","section":"Section 4.3, Eq. (1)-(2), Figure 5 (top)"},{"comment":"Baseline implementations are not specified in enough detail to verify the central comparison. The text says MVD, VIB, and Single Camera DrQ are reimplemented on top of the authors' DrQ baseline, and that MV-MWM is kept as a model-based baseline and 'tuned accordingly', but no hyperparameters, architecture details, loss weights, or training budgets are provided for these reimplementations. In addition, Section 5 says MV-MWM is trained 'without any expert demonstrations for a fair comparison', while Appendix A.2 says the MV-MWM baseline was kept as its model-based baseline; this should be clarified. Please provide full baseline configurations or a code release, because the reported 30% and 36% improvements over baselines are the paper's main empirical claim.","section":"Section 5, Baselines; Appendix A.2"}],"minor_comments":[{"comment":"The statement 'Qθ represents an exponential moving average of the weights from Qθ' is self-contradictory; please introduce explicit target-network notation (for example, Q-bar) and use it consistently in the critic target in Eq. (2).","section":"Section 3 and Eq. (2)"},{"comment":"The symbol alpha is used both for the SAC entropy temperature and for the MAD loss weight; this overloads the notation in Eq. (1)-(2) and Table 1. Please rename one of the two hyperparameters.","section":"Section 3 and Section 4.3"},{"comment":"The sentence 'Feature summation preserves the magnitude of different view features, such that the downstream actor and critic have a signal of how many views are inputted' is unclear, because summation makes the feature magnitude grow with the number of views rather than preserving a fixed magnitude. Please clarify the intended property.","section":"Section 4.1"},{"comment":"The observation shape is described as '(3×R(3×84×84))' in Section 5 but as 'Image Size (3, 84, 84)' in Table 1; please make the frame-stack and channel dimensions consistent across the two descriptions.","section":"Section 5 and Appendix A.1"},{"comment":"The final success values on the far-right 'Average' panels are very small and difficult to read in print; a table with the final mean and CI for each method and evaluation condition would improve reproducibility of the headline numbers.","section":"Figure 4 and Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The alpha tuning concern is the main reason for major revision: Figure 5 shows an extreme sensitivity around alpha=0.8, and the selection protocol is undisclosed. If the authors can demonstrate that alpha=0.8 is stable on held-out tasks or report complete tuning curves, the paper would be close to acceptable. The self-citation of SADA is not a hidden fit: alpha=0.5 explicitly recovers the prior objective, and the main benchmark comparisons are against external baselines. I would also ask the editor to ensure that the promised code release contains the baseline configurations, because the paper currently points to a project website rather than a permanent repository."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical RL paper with a genuinely new combination — feature summation plus SADA-style feature-level augmentation — and an unusually broad evaluation. The gains look real, but the alpha=0.8 choice is tuned on the same tasks used in the headline average, so the exact margins should be treated as conditional until the tuning protocol is disclosed.\n\nWhat's new: MVD disentangles, VIB merges, SADA augments images. MAD is the first to merge via summation and then apply the single-view features as augmentations to the actor and critic. The loss in Eqs. (1)-(2) is a clean extension of SADA with an alpha weight. The benchmark coverage is the biggest strength: 15 Meta-World and 5 ManiSkill3 tasks, 5 seeds, 95% CIs, ablations of components, alpha, merge methods, plus occlusion, modality, and scalability experiments. The central claim — better sample efficiency plus robustness to missing cameras — is supported.\n\nWhere it's soft: the stress-test note is right about alpha. Figure 5 shows alpha=0.8 at 56% vs alpha=0.5 at about 26% on the same 5 hard Meta-World tasks that appear in the main Meta-World average. The paper doesn't say how alpha was selected: no range, no seeds, no criterion. ManiSkill3's consistent improvement with the same alpha is reassuring, but it's a transfer, not a held-out tuning protocol. That makes the 30%/36% headline margins less crisp than they look.\n\nA second soft spot: the SADA recipe is imported from a self-cited prior paper, and part of the stabilization is attributed to that recipe. The \"MAD - Naive Both\" ablation uses an unweighted objective, so it can't separate the effect of the SADA-style target architecture from the effect of alpha weighting. That's a minor confound — the component ablations do show both merging and disentanglement contribute — but worth asking the authors to isolate. Baseline reimplementation details are also incomplete, which makes exact reproduction harder. The absence of real-robot validation is honestly acknowledged in the limitations.\n\nWho it's for: anyone working on multi-view or multi-camera RL for manipulation. It's a useful baseline and a practical method. I'd send it to peer review and ask for the alpha tuning protocol and baseline hyperparameters before accepting the numbers at face value.","headline":"Solid empirical multi-view RL paper; the merging-plus-feature-augmentation combo is new and the evaluation is broad, but the headline margins rest on an alpha tuned on the same tasks, so treat them as conditional.","tokens_in":19924,"tokens_out":2649,"would_cite":true,"duration_ms":24495,"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":"MAD trains multi-view policies that keep working when a camera goes missing.","keywords":["visual reinforcement learning","multi-view robot learning","feature disentanglement","data augmentation","robotic manipulation","sample efficiency","camera robustness","sensor failure"],"falsifier":"Run the MAD versus MAD-Naive-Both comparison from Figure 5 on all 20 tasks with a full hyperparameter sweep of $\\alpha$ and the same seed budget; if a tuned naive-both baseline reaches MAD's success on both merged and singular views, then the SADA-style target separation is not what carries the claimed robustness, and the central mechanism would be unsupported.","tokens_in":18832,"feed_emoji":"🤖","tokens_out":12104,"duration_ms":110067,"temperature":0.7,"pith_summary":"The central claim is that merging multiple camera views and disentangling them are complementary goals that a single training scheme can achieve. The proposed MAD algorithm encodes each camera view with a shared network, sums the view features into one merged representation for the actor and critic, and also feeds each single-view feature into those networks as a regularizing augmentation. The result, if the paper is right, is a policy that learns faster than methods that only merge or only disentangle views, and that can be deployed with any one camera. This matters because real robot deployments are burdened by multi-camera setups and are sensitive to camera failure. The paper reports higher success rates than four baselines across 20 simulated manipulation tasks, including on singular third-person views.","feed_headline":"Multi-view training makes robot policies survive camera loss","feed_subtitle":"Summing camera features during training teaches a policy to act on any single view at deployment.","key_machinery":"The mechanism is feature-level data augmentation built on feature summation. Each view $o_t^i$ is encoded by a shared CNN into $V_t^i$, and summation produces the merged representation $M_t = \\sum_i V_t^i$, which keeps the feature dimension identical for one or many views and preserves a signal of how many views contributed. The SADA stabilization recipe is then reused: in both actor and critic updates, targets are predicted only from the unaugmented merged stream $M_t$, while online predictions also see each single-view feature $V_t^i$ as an augmentation, and the two objectives are combined with hyperparameter $\\alpha$. This lets the downstream networks learn from both the merged state and every individual view without the unstable multi-state training that the paper's naive-both ablation exhibits.","core_discovery":"MAD's central claim is that merging and disentangling are complementary: training on the summed multi-view feature $M_t = \\sum_{i=1}^n V_t^i$ while regularizing the actor and critic with each single-view feature $V_t^i$ as a feature-level augmentation yields both higher sample efficiency and robustness to camera loss. The loss follows the SADA recipe, with learning targets computed only from the unaugmented merged stream while online Q-values and policy actions are also computed from each single-view stream, balanced by an $\\alpha$ weighting set to $0.8$. On 15 Meta-World and 5 ManiSkill3 tasks, the paper reports higher average success rates than MVD, VIB, MV-MWM, and single-camera DrQ across all-camera and singular-view evaluations, with the largest margins on the third-person views.","pith_inferences":["The paper leaves implicit that the recipe may generalize beyond cameras: any set of complementary features (modalities, proprioceptive channels, or subgoal encodings) could be summed and then re-fed individually as augmentations, turning MAD into a generic sensor-dropout training scheme.","Because feature summation preserves the magnitude of the summed features, the critic may implicitly learn to count the number of active views; testing whether value estimates shift with the number of summed features could reveal the mechanism behind the graceful degradation.","The authors note the absence of real-robot validation; a natural next test is sim-to-real transfer with one physical camera disabled, which would show whether the robustness survives calibration shifts and image noise.","The fixed $\\alpha=0.8$ suggests the merged stream should dominate but not exclude single views; a task-dependent or adaptive $\\alpha$ could improve robustness further on tasks where one view is much more informative than others."],"forward_implications":["A single MAD training run yields a policy deployable with any one camera, so the camera array can be simplified at deployment time without retraining or added parameters.","MAD does not depend on any particular view: in the occlusion experiment it keeps solving ManiSkill3 tasks when two cameras are pointed at uninformative scenes and only one view is useful.","The same merge-and-augment structure transfers across modalities: an RGBD-trained MAD policy can be evaluated with RGB-only or depth-only inputs, and its depth-only success rate exceeds a depth-only DrQ baseline.","The gains persist with more cameras: on five views in ManiSkill3, MAD's average success rate is above all baselines, and mixing different input resolutions costs little when images are resized to a common size."],"supporting_citations":[{"why":"Supplies the SADA stabilization recipe that MAD reuses to keep feature-level augmentation stable.","marker":"[8]"},{"why":"DrQ is the visual actor-critic backbone on which MAD and all reimplemented baselines are built.","marker":"[19]"},{"why":"MVD is the main disentanglement baseline; MAD aims to keep its view-reduction robustness while adding sample efficiency.","marker":"[7]"},{"why":"VIB is the main merging baseline; its first-person dependence motivates MAD's view-agnostic design.","marker":"[6]"},{"why":"MV-MWM is the multi-view world-model baseline that MAD is compared against on the same tasks.","marker":"[38]"},{"why":"Provides the alpha-weighted stabilization under augmentation from which MAD's loss weighting is borrowed.","marker":"[24]"}],"fun_headline_variants":["Merging and disentangling views yields efficient, robust robot policies","Multi-view training with disentanglement survives camera failure","Summing views for sample efficiency, separating for camera resilience","MAD: merge views to learn, disentangle to deploy robustly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the stabilization recipe designed for image-level augmentation also works when the augmented inputs are single-view feature vectors and the unaugmented input is their sum; if that transfer fails, the training degenerates into the unstable naive-both variant that the paper's own ablation reports.","fun_headline_variants_meta":{"raw":{"variants":["Merging and disentangling views yields efficient, robust robot policies","Multi-view training with disentanglement survives camera failure","Summing views for sample efficiency, separating for camera resilience","MAD: merge views to learn, disentangle to deploy robustly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00061,"raw_usage":{"total_tokens":2787,"prompt_tokens":844,"completion_tokens":1943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":1873}},"tokens_in":460,"tokens_out":1943,"duration_ms":15408,"temperature":1.0,"reasoning_tokens":1873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:23:46.241644+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the MAD versus MAD-Naive-Both comparison from Figure 5 on all 20 tasks with a full hyperparameter sweep of $\\alpha$ and the same seed budget; if a tuned naive-both baseline reaches MAD's success on both merged and singular views, then the SADA-style target separation is not what carries the claimed robustness, and the central mechanism would be unsupported.","supporting_citations":[{"cited_title":"Kostrikov, D","cited_arxiv_id":null,"evidence_quote":"DrQ is the visual actor-critic backbone on which MAD and all reimplemented baselines are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MV-MWM is the multi-view world-model baseline that MAD is compared against on the same tasks."},{"cited_title":"Hansen, H","cited_arxiv_id":null,"evidence_quote":"Provides the alpha-weighted stabilization under augmentation from which MAD's loss weighting is borrowed."}],"review_version":1}