{"id":"acf8da75-b440-4753-bd5d-fd540bf8c742","arxiv_id":"2610.17843","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PAF-Grasp forecasts future SE(3 object trajectories from short pose histories and selects reachable contact states, improving dynamic grasping success rates substantially over reactive baselines.","lead":"This paper proposes predicting only the future SE(3 pose of an object to guide a robot's grasping motion, rather than generating full videos or scene predictions. The key idea is that contact-relevant rigid-body state is a compact and actionable abstraction that enables real-time proactive dynamic grasping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Sim-to-real transfer robustness is unquantified; the inertial prior (I_B) is crucial for rotational accuracy (Table 4: Rot-FDE drops from 11.24° to 4.75° with I_B), but no experiment tests how VLM material-classification errors or mesh-reconstruction noise affect forecasting or grasping success.","rationale":"The reader's weakest_assumption flagged the sim-to-real transfer and noted that the paper does not quantify degradation in the inertial prior estimate. I agree that this is a key unaddressed issue. However, I consider it load-bearing because the success of the method hinges on the forecaster's ability to generalize across objects with different inertias, and the paper's own ablation shows I_B is critical for rotation. The proposed concrete test would either confirm robustness (strengthening the claim) or reveal a vulnerability (weakening it). The reader's verdict was CONDITIONAL already, and this concern adds a specific condition rather than overturning the existing judgment, so UNCHANGED is appropriate. The paper does provide other strong evidence: simulation comparisons with fair baselines (Table 1), NODE advantages over AR (Table 5, Fig. 7), and real-world results that, while comparing against a baseline trained on its own data, show large absolute margins. Thus the central claim remains plausible but requires the robustness check to become fully convincing.","tokens_in":20284,"tokens_out":3497,"duration_ms":30473,"concrete_test":"Run the toss forecasting experiment (Table 4 setting) with synthetic noise on I_B: scale the diagonal elements by factors 0.5, 0.75, 1.25, 1.5, and also add off-diagonal perturbations of up to 10% of the max moment. Record Pos-FDE, Rot-FDE, and the downstream success rate using the interception selector. If Rot-FDE stays below, say, 8° and success rate stays above 45% for all perturbations up to 25% error, the robustness is supported. If not, the sim-to-real transfer claim is materially weaker.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the forecaster transfers from simulated primitives to real objects via a recovered inertial prior is plausible but untested under realistic degradation. Section 3.3 states that the inertia prior ψ = I_B provides object-specific rotational structure, and Table 4 shows its large effect on toss Rot-FDE (11.24° → 4.75°). Yet the paper never injects noise into the prior in simulation to quantify robustness. The real-world pipeline (§F) uses a VLM to infer material category → density → inertia, and a single-view mesh recovry. Both sources can be significantly off e.g., wood vs. plastic density error >2×. Without an ablation, it is unknown whether the forecaster's learned acceleration field can compensate for such estmiation errors, or whether the real-world gains (67.3% vs. 15.3% depend on unusually accurate inertia estimates. The authors acknoledge 'coarse single-material density' and 'spatially varing mass distributions remain outside scope' but do not test how these limitions affect performance. This leaves a gap in the evidence for the sim-to-real transfer claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper proposes PAF-Grasp, a framework for proactive dynamic grasping that predicts future SE(3) rigid-body states from short pose histories (forecast) and then selects a reachable contact-time state under kinematic and timing constraints (select). The key insight is that contact-relevant SE(3) state is a compact, task-sufficient abstraction, avoiding the cost and indirectness of dense visual future prediction. A neural ODE forecaster is trained exclusively on simulated primitive shapes (spheres, cuboids, cylinders) with randomized dynamics, and transfers to real objects by recovering a metric mesh and an approximate inertia prior from a single RGB-D observation. Experiments across four simulated tasks and three real-world tasks report large success-rate improvements over reactive visuomotor baselines (e.g., average success from 24.5% to 62.0% in simulation, and from 15.3% to 67.3% in real-world), while running at 15 ms per inference.","tokens_in":20609,"tokens_out":4225,"duration_ms":35580,"significance":"If the claimed results hold, PAF-Grasp offers a practical and efficient approach for dynamic grasping by explicitly decoupling object-centric forecasting from robot-specific execution. The work provides systematic evidence that explicit physical state prediction outperforms dense visual world models in accuracy, physical consistency, and real-time efficiency for this contact-critical setting. The framework is lightweight, sim-to-real transferable with minimal object-specific setup, and compatible with real-time control loops. These findings support a broader shift from reactive policies and dense world models toward physically structured, action-relevant state abstractions for dynamic manipulation.","major_comments":[{"comment":"The inertia prior ψ=I_B is shown to reduce rotational final displacement error (Rot-FDE) from 11.24° to 4.75° on the toss task, indicating its crucial role for rotational forecasting. Yet the paper provides no quantitative analysis of how degradation in the estimated inertia prior (e.g., from VLM material classification errors, mesh reconstruction noise, or density estimation errors) affects forecasting accuracy or grasping success rate. The authors acknowledge in §5 that the inertia prior assumes a coarse single-material density but do not test robustness to realistic errors. Since real-world estimation of I_B is approximate, this omission weakens the evidence for the sim-to-real transfer claim. I recommend adding a simulation ablation that injects controlled noise into the inertia prior (e.g., scaling errors of ×0.5–×2, mass distribution misspecifications) and reports the impact on Rot-FDE, Pos-FDE, and success rate across motion families.","section":"§4.1, Real-world experiments","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-world comparison pits PAF-Grasp (which receives SE(3) pose histories, forecasts, and selected contact states) against π0.5 (which receives only RGB, language, and robot proprioception). The paper correctly states that this evaluates the effect of making contact states explicit, but the conflated factor of input modality prevents isolating the benefit of forecasting from the benefit of having object-centric SE(3) state inputs in general. A clearer ablation would be to evaluate a version of π0.5 augmented with current SE(3) poses (without forecasting), or to decompose the gain into the portions attributable to SE(3) input versus the forecast-then-select mechanism. This would clarify the core contribution of the forecasting component.","section":"§4.1, Simulation setup","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Simulation experiments use ground-truth object poses as forecaster inputs, which appropriately isolates forecasting performance. However, real-world results include perception errors from FoundationPose++ and mesh reconstruction. The paper does not quantify the sensitivity of overall system performance to perception noise. Adding a simulation ablation with realistic pose noise (based on measured perception accuracy) or reporting perception-noise statistics (e.g., translation/rotation RMSE of the tracker on the real setup) would improve the transparency of the sim-to-real transfer and help readers assess the robustness of the results.","section":"§4.1 and §F","authors_acknowledge":"In our real-world setup, high-speed 6D tracking [46] of the asymmetric hammer was sensitive to motion blur and rapid pose changes, making the measured success rate strongly coupled to percep- tion stability rather than only to the interception policy.","authors_acknowledge_p":0.75}],"minor_comments":[{"comment":"The abstract claims 'orders of magnitude faster than dense visual future prediction' but does not include the concrete runtime numbers for video models (30–180 s from Table 9) in the main text. Consider adding a brief runtime comparison in §4.2 to strengthen the point without relying on the appendix alone.","section":"Abstract and §4.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The success rate for VLASH on Conveyor is listed as '32' without a percent sign. For consistency, all entries in the SR columns should include the '%' symbol or be explicitly stated as percentages.","section":"Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The y-axis label on the left subfigure reads 'Position Error (m)' with an extra closing parenthesis in the PDF rendering. Please correct the label to 'Position Error (m)'. Also consider increasing font sizes in the legend for readability.","section":"Figure 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The quaternion derivative '˙˜q = 1/2 ˜q ⊗ [0, ω^B]' is given correctly, but the earlier description of the state vector could benefit from explicitly stating that the quaternion is stored as the relative rotation from the last observed orientation. A brief clarifying sentence immediately before the equation would reduce potential confusion.","section":"§3.2, Eq. (2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses the phrase 'coarse single-material density' in the limitations but does not give an estimated error range for typical VLM-based density estimation. While not required, a rough uncertainty quantification (e.g., density error ±x% from the VLM) would help readers gauge the realism of the prior.","section":"§3.3 and §5","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The core idea here is clean and sensible: instead of generating pixel-level video futures for dynamic grasping, predict the object’s future SE(3) state directly. That’s the right question to ask, and the authors make a good case for why it matters. The framework (PAF-Grasp) is clearly described, the sim-to-real recipe using primitive shapes plus an inertia prior is clever, and the latency numbers (15 ms) are genuinely impressive. The success rate gains—from 24.5% to 62% in simulation, 15.3% to 67.3% in the real world—are large enough to catch attention.\n\nNow the soft spots, and these are real. The comparison with video generation models is explicitly “not input-matched” ([b135]), which the authors acknowledge, but that makes the headline claim about “outperforming dense visual futures” weaker than it reads. The real-world baseline (π0.5) is trained on expert trajectories collected by the system itself, which conflates representation quality with data quality. There are no error bars reported anywhere, which matters when the absolute success rates are moderate. And the stress-test concern—that the inertia prior (I_B) is crucial for rotational accuracy (Table 4: Rot-FDE drops from 11.24° to 4.75° with I_B) yet the paper never tests how VLM material errors or mesh noise degrade forecasting—is a genuine gap. The authors list sensible limitations in the conclusion, but the central transfer claim would be stronger with a quantitative ablation on prior degradation.\n\nStill, this is a serious piece of work. The reasoning is sound, the experiments are systematic (if not perfectly controlled), and the argument that contact-relevant SE(3) is a better abstraction than pixel futures is worth taking seriously. The paper is for robotics researchers working on dynamic manipulation, time-critical grasping, or prediction-based control. I’d want to see the comparisons tightened and error bars added before fully endorsing the strongest claims, but the core contribution deserves a proper hearing.","headline":"A well-motivated paper arguing that compact SE(3) state prediction is the right abstraction for dynamic grasping, but the evaluation needs tighter baselines and robustness tests to fully back the claim.","tokens_in":21180,"tokens_out":1133,"duration_ms":12894},"pith_extraction":null,"created_at":"2026-10-14T14:41:10.707386+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}