{"id":"7722d106-d8d3-4ec3-ac0b-06a39ef12084","arxiv_id":"2511.12795","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ActiveGrasp selects the next camera view that maximizes predicted reduction in grasp-success entropy using a calibrated SE(3) energy-based model, and reports higher grasp success than prior active-grasping methods in simulation and real-robot tests.","lead":"This robotics paper proposes a next-best-view planner for grasping in clutter: it picks camera views that most reduce uncertainty about which grasp poses will succeed, using a calibrated energy-based model of grasp success on SE(3). A smart generalist should read it because it frames active perception around the grasp task itself instead of around scene coverage, and reports higher grasp success with fewer views in simulation and on a real robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Information-gain estimate rests on unvalidated Hessian approximation (Eq 16) and a possibly sign-flipped entropy definition (Eq 10); if either fails, selected views may not reduce grasp entropy, weakening the central claim.","rationale":"The reader's CONDITIONAL verdict is appropriate, and our strongest concern aligns with the reader's weakest assumption: the approximations underlying the information-gain estimate (Eq 16) are unvalidated and acknowledged as a limitation. However, we also identify a separate internal-consistency risk regarding the definition of grasp entropy in Eq 10, which appears to be the negative Shannon entropy. If taken literally, the sign of the information gain would be inverted, contradicting the claimed behavior in Fig. 3 and the maximization objective. This sign issue compounds the approximation concern because both affect the validity of the theoretical basis for view selection. Even if the empirical success rates are reproducible, they would not demonstrate that information gain is the driving factor unless the approximations and sign conventions are validated. Thus we recommend keeping the CONDITIONAL verdict, contingent on the authors releasing the appendix, code, and running the proposed validation of information-gain estimates. We do not see a need to move to REJECT or UNVERDICTED, as the empirical results remain plausible and the issues are addressable.","tokens_in":15935,"tokens_out":11079,"duration_ms":97868,"concrete_test":"Run a simulation study: select 10 random cluttered scenes from the evaluation suite. For each scene, sample the 128 candidate views and compute the approximate information gain I_approx via Eq 15/16. For a random subset of 20 candidates per scene, actually acquire each view (render and update 3DGS), recompute the calibrated EBM entropy η(w) after update, and measure the true information gain as H[g,w|D] − H[g,w|D,y] using the same η(w) (but with correct sign convention). Compute the Spearman rank correlation between I_approx and true information gain across candidates. If the correlation is below, say, 0.5, the approximation does not rank views correctly. Additionally, check whether the implementation of Eq 10 uses the literal negative entropy or the standard positive entropy; if literal, the algorithm's objective is sign-flipped. Either result would directly falsify the causal claim tha","verdict_should_be":"UNCHANGED","load_bearing_attack":"To justify that ActiveGrasp selects views by reducing grasp entropy, the paper must accurately estimate I in Eq 15. This requires: (a) the Hessian approximation ∇²_wη(w) ≈ ∇_wη(w)∇_wη(w)ᵀ + λI (Eq 16) to capture true curvature, and (b) the entropy definition Eq 10 to have the correct sign so that 'reduction' is positive when success becomes more certain. Neither is validated. Eq 16 is ad hoc and acknowledged as a limitation in Sec 5; no experiment compares predicted information gain to actual entropy reduction after acquiring a view. If the rank-one-plus-diagonal form misses off-diagonal curvature, views ranked as high-information may actually be uninformative, yet still outperform random due to other cues (e.g., viewing geometry), making the success-rate advantage coincidental rather than causal. Separately, Eq 10 as written gives h(s)=s ln s+(1−s)ln(1−s), the negative Shannon entropy. With this definition, η(w) is nonpositive and the quantity H[g,w|D]−H[g,w|D,y] would be negative when success probabilities move toward 0 or 1 (the behavior described in Fig. 3 as entropy decreasing). Thus the sign of I in Eq 15 is inconsistent with the stated objective unless Eq 10 is a typo for the standard entropy −s ln s−(1−s)ln(1−s). The paper's derivations of Eqs 14/15/17 are deferred to a missing appendix, so the reader cannot verify these sign conventions. These issues are internal-consistency risks, not merely disagreements with the consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ActiveGrasp, a next-best-view planning method for robotic grasping. It defines the information gain of a candidate view as the expected reduction in a grasp-specific entropy, where the grasp distribution is modeled by a calibrated energy-based model on SE(3). The scene is represented by 3D Gaussian Splatting, the posterior covariance is obtained via the Gaussian Approximation of the Posterior, and the Hessian of the grasp entropy is approximated by a rank-one-plus-diagonal form. The energy-based model is calibrated using failure grasps, an average-precision loss, and a learnable temperature. Experiments in a PyBullet benchmark and on a real robot report that ActiveGrasp achieves 79% grasp success with two actively selected views, versus 74.25% for the best baseline with the same calibrated grasp model and view budget. An ablation on ACRONYM and calibration metrics is also presented.","tokens_in":16420,"tokens_out":7156,"duration_ms":65339,"significance":"If the derivations are correct, the paper would make a useful contribution by replacing proxy objectives such as visibility or affordance with a task-specific information-gain estimate computed directly from a calibrated grasp distribution on SE(3). The paper has clear strengths: it addresses a well-motivated problem, provides a reproducible simulated benchmark (code is promised), includes a careful ablation of the calibration components, and the view-selection objective is not circularly fitted to the final success metric. However, the central information-gain derivation is deferred to an appendix that is not present in the submitted text, the key Hessian approximation is ad hoc and unvalidated, and the sign convention of the entropy definition appears internally inconsistent. These issues currently prevent full verification of the paper's central claim.","major_comments":[{"comment":"The entropy definition in Eq. (10) is h(s)=s ln s+(1-s) ln(1-s), which is the negative of the standard Shannon entropy of a Bernoulli variable. The text states that H[g|w] decreases when more successful grasps are discovered, but with Eq. (10), as s→1, h(s) increases from -ln2 toward 0. Moreover, if the standard entropy were intended (i.e., with a minus sign), then ∇²η(w) would be negative semidefinite, making Eq. (15) negative when evaluated as trace(∇²η (Σ0-Σ1)) with Σ0-Σ1 positive semidefinite. Under the definition as written, the quantity is not entropy and the claimed monotonic behavior in Fig. 3 is reversed. The two equations are mutually inconsistent, and this is load-bearing for the central claim that views are selected by reducing grasp entropy. Please correct the sign convention and re-derive Eqs. (14)-(15) consistently.","section":"Sec. 3.2, Eq. (10) and Eq. (15)"},{"comment":"The proofs of Eqs. (14), (15), and (17) are stated to be left to the appendix, but no appendix is included in the submitted manuscript. These equations are the core of the information-gain formulation, and the sign of Eq. (15), the Gaussian approximation in Eq. (14), and the derivative identity in Eq. (17) cannot be checked without the derivations. This is not a stylistic issue; it blocks verification of the central claim. The final version must include the full derivations or the paper should be reviewed with the appendix attached.","section":"Sec. 3.2, Eqs. (14), (15), (17)"},{"comment":"The approximation ∇²w η(w) ≈ ∇w η(w)∇w η(w)ᵀ + λI is introduced as 'working well in practice' but is not derived, compared with alternative approximations, or validated. The authors themselves list this as a limitation in Sec. 5. Since the entire view-selection signal is the trace of this Hessian against a covariance difference, an inaccurate Hessian can rank candidate views incorrectly. Please provide evidence that the predicted information gain correlates with the actual reduction in grasp entropy (or with grasp success) after acquiring a view, or replace Eq. (16) with a more principled approximation, e.g., a Gauss-Newton or Monte-Carlo estimate of ∇²η.","section":"Sec. 3.2, Eq. (16)"},{"comment":"The reported headline improvement (79% vs. 74.25%) is not accompanied by any statistical significance measure. With roughly 400 trials, the standard error of a binomial success rate is about 2 percentage points, so the difference is about 1.6 standard errors; the comparison with Random† (76.5%) is even weaker. The claim that ActiveGrasp outperforms the best baseline with the same view budget would be considerably strengthened by confidence intervals, a significance test, or per-scene/per-target variance reporting. As written, the 4.75-point gap may not be reliable.","section":"Sec. 4.3, Table 1"}],"minor_comments":[{"comment":"The '+2' in the denominator of the probability normalization is unexplained. It makes p_S + p_F < 1, leaving residual probability mass. Since calibration is central to the method, please justify this choice or provide an ablation showing its effect.","section":"Sec. 3.3, Eq. (19)"},{"comment":"The implementation details cite 'ACE [9]' for next-best-view selection, but the experiments and Table 1 use ACE with reference [71]. Please correct the citation to avoid ambiguity.","section":"Sec. 4.1 / References [9] and [71]"},{"comment":"The figure caption states that H is 'the grasp entropy defined in our paper,' but the plotted entropy curve is not labeled with the sign convention. Given the sign issue in Eq. (10), please make the plotted quantity explicit and ensure it matches the corrected definition.","section":"Fig. 3"},{"comment":"The model is variously called Se3diff, Se3diff Scene, Se3diff Calib, and Se3diff-Calib. Please unify the terminology across Table 1, Table 3, and the text.","section":"Naming"},{"comment":"The abstract and conclusion state that source code and the simulator benchmark will be released, but no link or release plan is given. For a reproducibility-oriented contribution, please provide a clear availability statement.","section":"Availability"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency between Eq. (10) and Eq. (15) and the missing appendix are the main blockers. If the sign is simply a typo and the appendix is provided, the framework may be sound, but the ad hoc Hessian approximation and the lack of statistical significance in Table 1 still need to be addressed. I do not see grounds for rejection, but the central claim cannot be accepted as currently written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading and worth refereeing. The core idea — phrasing next-best-view selection in terms of reducing the uncertainty of grasp success on SE(3), computed from a calibrated EBM — is genuinely new relative to visibility-, affordance-, or grid-entropy-based planners, and the experiments back it up: 79% vs 74.25% for the best baseline at the same view budget, and the ablation vs. random views shows the active selection contributes.\n\nThe parts that are solid: the calibrated EBM itself is a real improvement over SE(3)-Diffusion Fields, shown in Table 3 (AP up, ECE down). The simulation benchmark with YCB objects and 400 trials is a useful contribution, and they compare fairly against existing active methods using the same grasp model. The runtime is competitive.\n\nThe theory as written has three problems. First, Eq. 10 defines h(s) as s ln s + (1-s) ln(1-s), which is the negative Shannon entropy. That flips the sign of the information gain in Eq. 15 unless it's a typo — the accompanying text and Fig. 3 seem to expect the standard entropy with a minus sign. This needs to be fixed, not hand-waved. Second, the proofs of Eqs. 14, 15, and 17 are deferred to an appendix that isn't provided; since these are the core of the derivation, that's a blocker for verification. Third, the curvature approximation in Eq. 16 is ad hoc and acknowledged as a limitation; there's no validation that predicted information gain correlates with actual entropy reduction after a view. The unexplained '+2' in Eq. 19 is minor but should be clarified. The experiments would also be stronger with error bars or multiple seeds.\n\nThe central empirical claim holds up well enough to take seriously; the flaws are not fatal but they need addressing. The paper deserves a serious referee and, after the authors supply the missing appendix and fix the sign issue, probably acceptance with major revisions.","headline":"A serious, well-posed active grasping paper with a plausible central claim and real empirical gains, but the theoretical core currently rests on an unverified Hessian approximation, a missing appendix, and a likely sign error in the entropy definition.","tokens_in":16876,"tokens_out":3628,"would_cite":true,"duration_ms":30293,"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":"The next best view for grasping is the one that most reduces the entropy of the grasp-possibility distribution, not the one that covers the most scene.","keywords":["active grasping","next-best-view selection","energy-based model","SE(3) grasp poses","entropy reduction","model calibration","3D Gaussian Splatting","information gain"],"falsifier":"In a simulated scene with a small object and a dense candidate-view set, exhaustively evaluate every pair of additional views by physically executing a large set of grasps after each pair, and measure the true reduction in grasp-success entropy. If the pair ranked highest by ActiveGrasp's predicted entropy reduction does not yield the largest measured reduction within statistical error over repeated trials, the central claim is falsified.","tokens_in":15810,"feed_emoji":"🤖","tokens_out":10652,"duration_ms":82105,"temperature":0.7,"pith_summary":"ActiveGrasp aims to show that a robot's next view during grasping should be chosen by measuring how much that view would shrink the uncertainty in the grasp-possibility distribution, not by maximizing scene coverage or visibility. The paper defines the entropy of a grasp distribution as the expected entropy of individual grasp attempts, each modelled as a Bernoulli variable whose success probability comes from a calibrated energy-based model on the SE(3) pose manifold. In cluttered simulated scenes with two fixed initial views and two actively chosen views, the method achieves 79% grasp success, versus 74.25% for the strongest baseline planner running with the same calibrated grasp model. The paper also finds that eight additional random views cannot match the active four-view result, supporting the claim that where the robot looks matters more than how many views it takes. The calibration step—training the energy level against both successful and failed grasps with an average-precision loss and a learnable temperature—is what makes the entropy estimate meaningful.","feed_headline":"79% grasp success: entropy-guided views beat coverage","feed_subtitle":"Four actively chosen views reach 79% success; best coverage baseline gets 74%. Where to look beats how many views.","key_machinery":"The central object is a calibrated energy-based model of grasp poses on the SE(3) manifold (the six-degree-of-freedom space of positions and orientations). The model outputs separate energy values for success and failure for a scene represented by 3D Gaussian Splatting, and a learnable temperature; it is trained with an average-precision loss and on both successful and failed grasps so that the energy level aligns with success probability. The second piece is the entropy-reduction estimate: the expected information gain of a candidate view is expressed as half the trace of the curvature of the grasp-entropy function with respect to scene parameters, times the reduction in the posterior covar","core_discovery":"The central claim is that information gain for active grasping should be computed directly from the calibrated distribution of grasps, and that this is tractable. The authors define the entropy of the grasp distribution as η(w)=E_{p(g|w)}[h(s(g,w))], where h is the Bernoulli entropy of a single grasp's success probability s. Using a Gaussian Approximation of the Posterior of a 3D Gaussian Splatting scene, the expected entropy reduction from a candidate view becomes half the trace of the curvature of η with respect to scene parameters, times the reduction in scene-parameter covariance. Because the energy model alone gives only relative scores, the paper calibrates its energy to success rate b","pith_inferences":["A natural extension: the same entropy-reduction objective could choose views for other manipulation skills, such as tool use or assembly, whenever an energy-based model supplies a distribution over the relevant action manifold.","The rank-one-plus-diagonal curvature approximation is the piece most likely to fail under domain shift or with richer scene representations; a stochastic or block-diagonal Hessian estimator is a direct testable improvement.","A consequence the authors do not pursue: if view placement matters more than view count, a perception system with a tight view budget could still succeed by spending all of it on entropy-optimal viewpoints, relevant for low-power or time-critical robots.","Because calibration was performed on simulated data, an online recalibration loop using real executed grasps would test whether the calibration guarantee survives the sim-to-real gap the authors acknowledge."],"forward_implications":["Using two fixed and two actively chosen views, ActiveGrasp reaches 79% grasp success in simulated clutter, against 74.25% for the best baseline planner with the same calibrated grasp model.","Calibration matters: pairing the same entropy-driven view selection with an uncalibrated energy model yields 73%, and the calibrated model cuts expected calibration error on executed grasps from 0.35 to 0.02.","Ten random views (eight more than active methods) reach only 76.5%, implying that view placement, not view count, is the dominant factor in grasp success.","Because information gain is defined directly from the grasp distribution, the approach removes the need for separate affordance predictors, coverage heuristics, or other surrogate rewards.","In three real-world cluttered scenes, the method succeeds on 9/10, 8/10, and 9/10 grasp attempts, outperforming coverage-, affordance-, and graspness-based planners that share the same calibrated grasp model."],"fun_headline_variants":["Calibrated grasp energy guides active views to 79% success","Entropy of grasp distribution picks better views","Active grasping with calibrated energy models beats coverage","Information-guided views outperform coverage in cluttered grasps","Grasp-distribution entropy drives active view selection"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The view-quality ranking stands or falls on two approximations: the scene posterior being Gaussian with an inverse-Hessian approximated as diagonal, and the curvature of grasp entropy being the gradient outer product plus a diagonal term; if either misestimates the true entropy reduction, the selected views may not actually be the information-optimal ones.","fun_headline_variants_meta":{"raw":{"variants":["Calibrated grasp energy guides active views to 79% success","Entropy of grasp distribution picks better views","Active grasping with calibrated energy models beats coverage","Information-guided views outperform coverage in cluttered grasps","Grasp-distribution entropy drives active view selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1541,"prompt_tokens":759,"completion_tokens":782,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":503,"tokens_out":782,"duration_ms":5876,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:56:59.685018+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a simulated scene with a small object and a dense candidate-view set, exhaustively evaluate every pair of additional views by physically executing a large set of grasps after each pair, and measure the true reduction in grasp-success entropy. If the pair ranked highest by ActiveGrasp's predicted entropy reduction does not yield the largest measured reduction within statistical error over repeated trials, the central claim is falsified.","supporting_citations":[],"review_version":1}