{"id":"7bef9c88-4853-4453-a1fb-343fe235a7f9","arxiv_id":"1909.02129","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A simulation-trained pair of CNNs predicts the distribution of post-grasp object displacement, and selecting low-variance grasps reduces real-world placement errors to 0.68 cm and 3.42 degrees.","lead":"This paper trains two neural networks in simulation to predict how a robot's grasp will shift an object, then uses the predicted shift to plan more precise grasps. It reports centimeter-level placement accuracy on a real robot without retraining on real data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world validation uses only planar tape-marker measurements, so the 0.68 cm / 3.42 deg headline ignores Δz and out-of-plane tilt; the 'precise grasping' claim is narrower than stated until full 6-DOF pose error is measured.","rationale":"The reader's weakest assumption identifies the most load-bearing concern: the real-world metric that supports the central claim is computed under an unvalidated planar-motion assumption. The paper has genuine independent support: a large simulation dataset with object-wise splits, sim-to-real deployment without fine-tuning, consistent comparison among GDN variants, and reduced error bars when selecting low-variance grasps. None of these, however, repairs the fact that the headline 0.68 cm / 3.42 deg is a 2-D/1-D measurement and the method's stated output includes Δz. Because the central claim explicitly says 'precise robotic grasping' and downstream compensation, the missing full-pose validation is the condition on which the claim should be accepted. The reader's CONDITIONAL verdict already reflects this, so no verdict adjustment is needed. If the proposed motion-capture test later shows negligible z and tilt errors, the condition would be satisfied; if not, the claim should be narrowed to planar placement precision.","tokens_in":8644,"tokens_out":11282,"duration_ms":121108,"concrete_test":"Run the same real-robot trials on the same 7 objects with a rigid 6-DOF motion-capture marker cluster or an independent high-resolution stereo/structured-light sensor, and log full object pose at three instants: pre-grasp, immediately after lift, and after returning to the grasp pose. Compute RMSE separately for Δz, roll, pitch, and yaw, and compare with the reported 0.68 cm / 3.42 deg planar values. If Δz or out-of-plane tilt RMSE is comparable to or larger than the planar errors, the headline understates the true displacement; if they are negligible, the planar-only validation would be sufficient. Also check whether pose changes between lift and return are negligible to rule out a transport-slip confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim rests on a planar-only measurement. Section 3.1 defines the post-grasp displacement as [Δx, Δy, Δz, Δθ]^T, and the GDN is trained to predict all four components, including Δz and its variance. Section 4.4, however, measures real-world displacement with two top-mounted tape markers, which can recover at most x, y, and planar yaw; Δz and out-of-plane tilt are unobserved. The reported translation RMSE of 0.68 cm is therefore a 2-D in-plane error, not a 3-D translation error, and the 3.42 deg rotational RMSE is yaw-only. For palletizing or assembly, vertical placement and out-of-plane orientation matter; if the gripper lifts objects with variable z (varying seating depth in the jaws) or permits roll/pitch, the reported errors understate the true post-grasp pose uncertainty. The problem statement itself includes Δz, so the modeled quantity is not the measured quantity in the real-world experiment. This gap is load-bearing because the headline 'precise robotic grasping' and the claimed compensation for downstream tasks are validated only for planar placement until full 6-DOF pose error is assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of planning robotic grasps that yield a known and controllable post-grasp object pose. It trains two convolutional networks in simulation from 1011 CAD models: a Grasp Quality Network (GQN) predicting lift success, and a Grasp Displacement Network (GDN) predicting a Gaussian distribution over post-grasp displacement (mean and variance) from a depth image and grasp parameters. A planner filters grasps by predicted quality (top 3%) and then selects the grasp with the lowest predicted displacement variance; the predicted mean displacement is used to compensate downstream placement. The method is evaluated in simulation against LOWESS and ablated GDN variants, and on a real Franka Panda with seven novel 3D-printed objects. Headline results are a real-world translation RMSE of 0.68 cm and rotation RMSE of 3.42 deg for the GCIP-M+V GDN under high-quality, low-variance grasp selection.","tokens_in":8896,"tokens_out":6103,"duration_ms":55593,"significance":"If the results hold, the paper makes a useful contribution: it reformulates precision grasping as probabilistic displacement prediction, uses a correct heteroscedastic regression loss, and gives evidence that selecting low-variance grasps improves prediction consistency. The depth-only, simulation-to-real deployment without fine-tuning is a strength, and the paper states that datasets and supplementary material are available, which supports reproducibility. However, the evaluation as presented establishes only planar (2-D plus yaw) pose accuracy in the real world, and the simulation numbers mix training and validation objects; the central 'precise grasping' claim is therefore narrower than stated. With additional experiments quantifying full 6-DOF post-grasp error and a comparison to an existing precision-grasping baseline, the contribution would be solid.","major_comments":[{"comment":"The real-world evaluation measures only planar displacement. The problem statement defines Δp = [Δx, Δy, Δz, Δθ]^T and Eqs. (1)-(2) model a distribution over all four components, but §4.4 estimates displacement from two tape markers on top of the object, which can recover at most (Δx, Δy, Δθ) under the stated assumption that displacements occur only in a plane. Therefore the reported 0.68 cm translation RMSE and 3.42° rotation RMSE are in-plane errors, not 3-D pose errors; Δz and out-of-plane tilt are unmeasured. Since the motivating applications are palletizing and assembly, for which vertical placement and out-of-plane orientation matter, the headline claim of 'precise grasping' is not yet supported. Please report full 6-DOF post-grasp pose error (e.g., with an external tracker or CAD-model registration with symmetry handling), or at least provide an estimate or bound on Δz, and re-state the claims to the planar setting if this is not possible.","section":"§3.1, §4.4"},{"comment":"The simulation evaluation of the GDN uses 85 objects, of which 50 come from the training dataset and 35 from the validation set. Aggregating RMSE over this mixture can mask the method's performance on genuinely novel objects, especially since the planner's grasp selection is also learned. The paper should report separate results for training, validation, and held-out objects, and the headline simulation numbers should be based only on objects not seen during training or model selection. The real-world experiment already uses seven novel objects, but it is subject to the planar-measurement limitation in Major Comment 1.","section":"§4.3"},{"comment":"The palletizing application is demonstrated qualitatively with a single image and no quantitative placement error. To support the claim that the predicted displacement can be used to compensate for downstream tasks, the paper should measure the final placed-pose error (translation and rotation) for compensated versus uncompensated placements over multiple trials. Without this, the application-level benefit of the low-variance grasp selection is not established.","section":"§4.4.1, Fig. 8"},{"comment":"The baseline comparison is insufficient to support the claim of a precise-grasping contribution. LOWESS is limited to known objects and cannot generalize, and the remaining comparisons are ablations of the proposed method; no existing precision-grasping method (e.g., the probabilistic analytical approach of Chen et al. [4], or a Dex-Net-based planner [17]) is compared. A comparison to at least one prior precision-oriented grasp planner is needed to substantiate the 'precise robotic grasping' claim in the title and abstract.","section":"§4.1"}],"minor_comments":[{"comment":"The abstract says 'mean prediction errors of 0.68cm and 3.42deg' but the body reports root mean square error; please use consistent terminology.","section":"Abstract"},{"comment":"The factor of 1/2 in Eq. (4) is dropped in Eq. (5); this is a constant scaling and does not affect the optimum, but the two equations should be reconciled so that the loss is written consistently.","section":"Eqs. (4)-(5)"},{"comment":"There is a duplicated word in the sentence 'The robot robot executes a grasp g'.","section":"§3.3"},{"comment":"The sentence introducing the palletizing example refers to 'Figure 7', but the palletizing application is shown in Figure 8.","section":"§4.4.1"},{"comment":"Please state the units of the reported RMSE values in the axis labels or captions, and consider reporting per-object error distributions or a paired significance test, since the number of real-world trials is modest.","section":"Figs. 4-7"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of circularity; the main risk is evaluation completeness rather than derivational correctness. The paper's structure and evaluation depth are closer to a workshop paper than a full archival article. For a full journal or conference version, the planar-measurement gap, the training-object contamination in the simulation evaluation, and the missing precision-grasping baseline should be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core take: the paper does something new and useful. It trains a CNN to predict a Gaussian distribution over post-grasp displacement and uses the predicted variance to select low-variance, high-quality grasps. The entire pipeline is trained in simulation and deployed on a real Franka Panda without fine-tuning, which is a genuine practical achievement. The real-world numbers (0.68 cm translation, 3.42 deg rotation) are the headline, but they only describe in-plane translation and yaw rotation. The paper's own problem statement includes Δz, so the measured quantity in the real world is not the modeled quantity. The authors are transparent about the planar assumption, but the conclusion 'precise robotic grasping' is broader than the evidence supports. That is the paper's real soft spot, and it is not minor for palletizing or assembly where vertical placement and out-of-plane tilt matter. What the paper does well: the MLE loss is correct, the architecture choices are sensible, and the sim dataset is large (over 1M grasp attempts across 1011 CAD models). The observation that including variance prediction improves the mean prediction is a nice empirical result, and the comparison among full-image vs grasp-centric patches and mean-only vs mean+variance is clearly organized. Sim-to-real transfer without fine-tuning works, which is the strongest practical claim. Other issues, in proportion: the simulation evaluation of the GDN includes 50 objects from the training set among 85 total, though 35 held-out objects are also reported. The object filtering by grasp difficulty is post hoc and narrows the evaluation set. The real-world evaluation uses only seven objects, so generalization is thin. There is also no comparison against prior precision-grasping baselines like Chen et al., so we don't know the competitive position. These are real gaps but not a broken core. The method is clearly described, the math checks out, and the planar limitation is a scope restriction rather than a logical flaw. This paper is for people working on task-oriented grasping, grasp planning under uncertainty, and depth-only sim-to-real transfer. It is not a paradigm shift, but it is a sound incremental contribution and the variance-based grasp selection is a useful idea. A serious referee should engage with it; a good review would push for a full 6-DOF real-world evaluation and at least one baseline comparison. That would turn a plausible planar result into a convincingly precise one.","headline":"A sound, incremental contribution on probabilistic post-grasp displacement prediction that deserves a serious referee, but the headline real-world numbers only measure planar motion, so the 'precise grasping' claim is narrower than stated.","tokens_in":9429,"tokens_out":2219,"would_cite":true,"duration_ms":25235,"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":"A robot can grasp precisely by predicting how the object will shift during the grasp, then compensating.","keywords":["precise grasping","post-grasp displacement","grasp quality","variance prediction","sim-to-real","depth images","industrial parts","Gaussian displacement model"],"falsifier":"Repeat the experiment with a motion-capture or depth-based 6-DOF tracker while the object is grasped. If the predicted planar displacement corrections miss measured out-of-plane motion or the full-pose error exceeds the reported 0.68 cm and 3.42 degrees, the central claim is limited to planar cases.","tokens_in":8437,"feed_emoji":"🤖","tokens_out":7886,"duration_ms":77988,"temperature":0.7,"pith_summary":"The paper proposes that precise robotic grasping does not require eliminating object motion during a grasp; it requires predicting that motion. The authors train two convolutional networks in simulation on over a thousand industrial CAD parts: one predicts the chance a grasp lifts the object, the other predicts the mean and variance of the object's post-grasp displacement from a depth image and grasp parameters. On a real robot with novel 3D-printed parts, the best displacement model reaches a mean translational error of 0.68 cm and rotational error of 3.42 degrees, and selecting low-variance grasps makes those errors more consistent. The practical stake is that assembly and palletizing tasks can use the predicted mean displacement as a correction instead of relying on a rigid or precisely repeatable grasp.","feed_headline":"Grasps land within 0.68 cm by predicting object shift","feed_subtitle":"A simulation-trained network predicts post-grasp shifts, letting the robot pick low-variance grasps for accurate placement.","key_machinery":"The load-bearing object is the Grasp Displacement Network (GDN), a convolutional network that maps a depth image cropped to the grasp and aligned with its rotation, plus the grasp's offset from the object center, to a four-dimensional Gaussian over post-grasp displacement in the object frame. It is trained with a negative log-likelihood loss that includes the variance outputs, which lets the network raise its uncertainty on inherently variable grasps instead of fitting a single mean. A companion Grasp Quality Network, trained first and used to initialize the GDN's convolutional filters, restricts planning to the top 3% of grasps by predicted lift success; the planner then chooses the lowest-variance grasp among those. This two-stage selection is what turns a displacement predictor into a precise-grasp planner.","core_discovery":"The paper's central claim is that post-grasp displacement is a learnable random variable: for a top-down parallel-jaw grasp on a flat surface, the displacement of the object can be modeled as a Gaussian whose mean and diagonal variance are functions of the grasp and a 64x64 depth image. The Grasp Displacement Network learns these functions by maximizing the log-likelihood of simulated successful grasps, and the planner filters for the top 3% of grasps by predicted lift success, then picks the one with the smallest predicted displacement variance. The prediction is converted from grasp-frame to object-frame displacement and used as a compensation when placing the object. The reported real-world results, 0.68 cm and 3.42 degrees RMSE on novel objects, are the evidence that this distribution-based strategy supports precise placement.","pith_inferences":["The tape-marker evaluation assumes planar motion, so the paper's precision claim is directly established only for objects that slide without lifting or tilting; a 6-DOF motion-capture test would show how far the method generalizes.","Predicted variance can be read as a learned measure of grasp repeatability, which might also guide viewpoint selection or trigger a re-grasp when no low-variance grasp exists.","The same probabilistic displacement loss could be applied to other manipulation outcomes, such as pushing or insertion corrections, whenever a planner needs to select actions with predictable consequences."],"forward_implications":["Precise placement no longer requires a grasp that holds the object rigidly; a predictable shift can be corrected at placement time.","In real-world trials, planning for high quality and low variance reduced translation RMSE from 0.72 to 0.68 cm and rotation RMSE from 3.79 to 3.42 degrees while shrinking the spread across objects.","Simulation-only training on depth images transfers to a real robot without fine-tuning, so the pipeline could be applied to new grippers or depth sensors by regenerating simulated grasps.","The predicted mean displacement can offset the target pose in applications such as palletizing, where the robot must place objects at a known position and orientation."],"supporting_citations":[{"why":"Supplies the simulation-trained CNN grasp-quality approach from depth data that the GQN is built on.","marker":"[17]"},{"why":"Defines the grasp-quality formulation and task-oriented self-supervised grasping in simulation used here.","marker":"[9]"},{"why":"Shows depth-only 6-DOF grasping that transfers from simulation without fine-tuning, motivating the deployment strategy.","marker":"[10]"},{"why":"A prior precision-grasping method that minimizes post-grasp displacement, the contrast to this paper's predict-and-compensate approach.","marker":"[4]"},{"why":"Provides the simulation environment used to generate the 1.011 million grasp attempts for training.","marker":"[20]"},{"why":"The locally weighted regression baseline for displacement estimation on known objects.","marker":"[6]"}],"fun_headline_variants":["Predicting object shift yields grasps precise to 0.68 cm","Probabilistic post-grasp estimation boosts robotic precision","Sim-trained CNNs forecast grasp displacement for accuracy","0.68 cm accuracy by modeling post-grasp movement","Distribution-aware grasping achieves sub-cm placement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world precision numbers assume the object moves only in the table plane, and the tape-marker measurement leaves out vertical motion and tilts, so the method is validated for planar placement rather than general 6-DOF displacement.","fun_headline_variants_meta":{"raw":{"variants":["Predicting object shift yields grasps precise to 0.68 cm","Probabilistic post-grasp estimation boosts robotic precision","Sim-trained CNNs forecast grasp displacement for accuracy","0.68 cm accuracy by modeling post-grasp movement","Distribution-aware grasping achieves sub-cm placement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3660,"prompt_tokens":854,"completion_tokens":2806,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2726}},"tokens_in":470,"tokens_out":2806,"duration_ms":19074,"temperature":1.0,"reasoning_tokens":2726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:58:23.267807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the experiment with a motion-capture or depth-based 6-DOF tracker while the object is grasped. If the predicted planar displacement corrections miss measured out-of-plane motion or the full-pose error exceeds the reported 0.68 cm and 3.42 degrees, the central claim is limited to planar cases.","supporting_citations":[{"cited_title":"Robotics: Science and Systems (RSS)","cited_arxiv_id":null,"evidence_quote":"Supplies the simulation-trained CNN grasp-quality approach from depth data that the GQN is built on."},{"cited_title":"Robotics: Science and Systems (RSS)","cited_arxiv_id":null,"evidence_quote":"Defines the grasp-quality formulation and task-oriented self-supervised grasping in simulation used here."},{"cited_title":"Conference on Robot Learning (CoRL) 14 Jialiang (Alan) Zhao, Jacky Liang, and Oliver Kroemer","cited_arxiv_id":null,"evidence_quote":"Shows depth-only 6-DOF grasping that transfers from simulation without fine-tuning, motivating the deployment strategy."},{"cited_title":"Journal of Intelligent & Robotic Systems 90(1-2):19–43","cited_arxiv_id":null,"evidence_quote":"A prior precision-grasping method that minimizes post-grasp displacement, the contrast to this paper's predict-and-compensate approach."},{"cited_title":"In: 2013 IEEE/RSJ International Conference on Intel- ligent Robots and Systems, IEEE, pp 1321–1326","cited_arxiv_id":null,"evidence_quote":"Provides the simulation environment used to generate the 1.011 million grasp attempts for training."},{"cited_title":"Journal of the American statistical association 74(368):829–836","cited_arxiv_id":null,"evidence_quote":"The locally weighted regression baseline for displacement estimation on known objects."}],"review_version":1}