{"id":"7028ec2b-9399-42e6-b637-d3cfb106dd15","arxiv_id":"2509.07162","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Planning trajectories to many grasp targets before evaluating them, and scoring the resulting terminal grasps, improves multi-finger grasping success over the standard generate-evaluate-plan order.","lead":"This paper proposes a grasping pipeline that first plans robot trajectories to many candidate grasps in parallel, then uses a learned evaluator to score the grasp actually reached by each trajectory, and executes the most promising one. The method reports higher grasp success than the traditional evaluate-then-plan pipeline in simulation and on a real robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FPTE's central claim depends on the evaluator grading off-target terminal grasps, but the supporting PR curve is filtered after approach collisions and the evaluated grasps pair terminal poses with target finger configurations not seen in training; this transfer is not established.","rationale":"The reader's conditional verdict is appropriate. The paper has real strengths: the sim experiments cover three generators, two motion planners, and 20 training objects, and the real-world evaluation uses 11 novel objects, so the ordering effect is not a single-configuration artifact. However, the mechanism that makes FPTE work is precisely that the evaluator can score terminal configurations that are far from the target and in environments never seen during training. The paper's own Figure 8 PR curve is filtered after approach collisions, which removes a central failure class for off-target trajectories, and the runtime grasp construction pairs target theta_g with off-target terminal poses, a combination absent from training. These are not fatal because the reported success rates are unfiltered, but they mean the central claim is not fully established. A baseline matched in number of planning attempts and an unfiltered PR curve would close the gap. I agree with the reader's weakest assumption and keep the conditional verdict.","tokens_in":13030,"tokens_out":12288,"duration_ms":117107,"concrete_test":"Recompute the evaluator's precision-recall curve on exactly the set of trajectories FPTE ranks in the real-world experiments, without the Section IV.B filter that removes approach-collision samples, and stratify the curve by whether the terminal configuration reached the target. If the unfiltered curve, or the off-target subset, is materially worse than the reported Figure 8, the evaluator cannot be trusted to grade the off-target terminal states on which FPTE relies, and the claimed advantage would need an alternative explanation beyond plan-then-evaluate ordering.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FPTE removes the target-distance criterion and selects the trajectory whose terminal grasp is scored highest by the learned evaluator. The evaluator must therefore be reliable on terminal configurations that do not reach the proposed target. Two properties of the paper make this the load-bearing weak spot. First, Section III.A constructs the scored grasp by pairing the terminal pose and pre-grasp configuration from the planned trajectory with the in-grasp configuration theta_g taken from the original target grasp, whereas Section III.B collects training data by planning theta_g for the actual terminal configuration via contact optimization. The exact (pose, theta_p, theta_g) triples that FPTE ranks, especially for off-target trajectories, are out of the training distribution. Second, Figure 8, the only direct evidence that the evaluator transfers to the real world, is explicitly computed after filtering out samples where the robot hit the object during approach. Collisions during approach are a likely failure mode of off-target trajectories, so the filtered PR curve can overstate the evaluator's ability to rank the full candidate set. The 80% versus 22% real-world comparison is further confounded by the baseline being capped at three planning attempts while FPTE plans to 512 targets, though the reported 38% executed-success rate for the baseline partially mitigates this. The central mechanism is plausible, but its key component, the evaluator's off-target terminal-state scoring, is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FPTE (First Plan, Then Evaluate), a dexterous grasping pipeline that first uses a generator to propose a batch of target grasps, plans trajectories to all targets in parallel with a vectorized motion planner, and then uses a learned evaluator to score the grasp at each trajectory's terminal configuration, executing the highest-scoring trajectory. This reverses the standard generator-evaluator-planner order and explicitly disregards distance to the planned target. The authors report consistent simulation improvements across three generators and two motion planners, and a real-world success rate of 80% versus 22% for the baseline on 11 novel objects, with demonstrations in cluttered scenes and varied table/shelf heights.","tokens_in":13315,"tokens_out":4250,"duration_ms":38655,"significance":"If the claimed results hold, FPTE is a practically valuable reordering of a common grasping pipeline and a compelling application of vectorized planning, with a plausible explanation for why target-distance rejection discards successful grasps. The paper's strengths include a large simulated dataset (28.9M grasp attempts), systematic variation of generators and planners, and real-world generalization to novel objects and environments. However, the central evidence depends on unquantified out-of-distribution evaluator inputs and on a real-world comparison with asymmetric planning budgets; these issues need to be addressed before the quantitative claims are fully supported.","major_comments":[{"comment":"At runtime, the evaluator receives grasp triples (T_OH, theta_p, theta_g) in which the terminal pose and pre-grasp configuration come from the planned trajectory, but theta_g is taken from the original target grasp. In training, theta_g is planned for the actual terminal pre-grasp configuration via contact optimization, as described in Section III.B. The exact inputs that FPTE ranks, especially for off-target trajectories, are therefore out of the training distribution. Because FPTE's core claim is that the evaluator can grade these terminal states without the target-distance criterion, the paper should provide direct evidence of this transfer, such as a distributional comparison, an unfiltered evaluation set, or retraining/conditioning the evaluator with candidate-specific theta_g values.","section":"Section III.A and III.B"},{"comment":"The real-world precision-recall curve is computed after filtering out samples where the robot hit the object during the approach to the pre-grasp configuration. Collisions during approach are a natural failure mode for off-target trajectories, so filtering them can materially inflate the apparent quality of the evaluator's ranking on the full candidate set. Please report the PR curve on unfiltered data and, if possible, separately for trajectories that reach the target versus those that do not.","section":"Section IV.B, Figure 8"},{"comment":"The headline real-world comparison of 80% versus 22% is confounded: the baseline is capped at three planning attempts, while FPTE plans to 512 targets in parallel and then executes the best one. The 23 out of 55 baseline failures to find a trajectory may be partly an artifact of this budget. Even the executed-only 38% success rate lacks confidence intervals or a per-object/per-location breakdown, so the magnitude of the improvement is not established. Please report per-condition counts with binomial confidence intervals, or run a baseline with a comparable planning budget.","section":"Section IV.B"}],"minor_comments":[{"comment":"The simulation results in Figure 4 are reported without error bars or the number of repeated trials per condition; adding trial counts and variance estimates would make the cross-generator and cross-planner claims easier to assess.","section":"Section IV.A, Figure 4"},{"comment":"The sentence \"We model all the detected objects and the environment (hand-coded) as collisions for the motion planner\" appears verbatim twice in Section IV.B; one occurrence should be removed.","section":"Section IV.B"},{"comment":"The phrase \"with the promptobject that can be picked up with one hand\" appears to contain a missing-space typo; please correct it to \"with the prompt 'object that can be picked up with one hand'.\"","section":"Section IV.B"},{"comment":"The notation for resulting grasps G_R_OH and target grasps G_T_OH is dense; a short symbol table or a clarifying sentence would improve readability.","section":"Section III.A"},{"comment":"The limitations paragraph names object-shape diversity and segmentation inaccuracy but does not acknowledge the evaluator distribution shift discussed above; this should be added to the limitations.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"No concerns about novelty or citation practices. The main risk is statistical and distributional, as detailed in the report; the paper's central idea is plausible but the current evidence does not yet fully establish the evaluator's reliability on off-target terminal configurations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: FPTE delivers a simple, useful inversion of the standard generator-evaluator-planner pipeline, and the empirical win is large enough to take seriously. The catch is that the paper's key mechanism—the evaluator scoring off-target terminal grasps—has a distribution mismatch and a filtered PR curve, so I'd want revisions before treating the real-world numbers as fully established.\n\nWhat's new: planning to all proposed grasps in parallel with a vectorized planner and then evaluating the resulting terminal configurations, discarding the distance-to-target acceptance threshold. That's a clean idea, and it's well-motivated: the paper shows only 3% of its successful grasps would survive the default Curobo threshold. The ablations across three generators and two planners are real work, and the 28.9M full-robot simulation dataset is a substantial engineering effort. The real-world 80% vs 22% on 55 trials is a big difference, and the 38% executed-only baseline rate shows the effect isn't purely about FPTE getting more attempts.\n\nSoft spots, in order of importance. (1) The evaluator is trained on grasps where the in-grasp configuration θ_g is contact-optimized for the terminal pose, but at runtime FPTE pairs the terminal pose with the θ_g from the original target. Off-target terminal grasps are exactly the ones the approach needs to score, and those (pose, θ_p, θ_g) triples are out of training distribution. The paper doesn't address this mismatch. It's not fatal—the empirical results suggest the evaluator generalizes—but it's the load-bearing assumption and needs a dedicated analysis. (2) Figure 8, the only direct evaluator-transfer evidence, is computed after filtering out approach collisions. That filter is reported in the text, but it's easy to miss, and approach collisions are a likely failure mode for off-target trajectories. Show the unfiltered curve. (3) The real-world baseline is capped at three attempts while FPTE plans 512 trajectories at once. The 38% executed-success rate helps, but an explicit matched-attempt experiment or a clear acknowledgment of the confound is needed. Minor: no error bars or significance tests anywhere; with n=55 the headline gap is probably real, but sim results should show variance.\n\nBottom line: this is a solid extension-level contribution with a genuinely useful ordering idea. The evaluation is strong enough to deserve a serious referee, but I'd recommend conditional acceptance with the evaluator-mismatch and PR-curve issues addressed. I'd bring it to reading group and would likely cite it when working on grasping pipelines.","headline":"A clean plan-then-evaluate inversion with a large empirical win, but the evaluator's off-target scoring is under-validated and the real-world comparison needs matched attempts.","tokens_in":13824,"tokens_out":2906,"would_cite":true,"duration_ms":25097,"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":"The paper claims that planning to many grasp targets first and then scoring the terminal poses of the planned trajectories raises grasp success over the standard generate-score-plan order across generators, planners, and real-world scenes.","keywords":["robot grasping","multi-fingered hand","grasp success evaluation","vectorized motion planning","generate-evaluate-plan","sim-to-real transfer","point cloud grasping","parallel trajectory optimization"],"falsifier":"Run FPTE in simulation with the goal tolerance deliberately widened so many planned trajectories terminate far from their targets, then compare evaluator scores against actual lift success separately for the off-target subset, without filtering out approach collisions; if the evaluator is no better than chance on that subset, the claimed advantage of terminal-configuration scoring collapses.","tokens_in":12849,"feed_emoji":"🤖","tokens_out":10353,"duration_ms":86362,"temperature":0.7,"pith_summary":"This paper is trying to establish a simple reordering for learning-based robot grasping: instead of generating grasp candidates, scoring them, and then planning a trajectory to the best-scoring one, a robot should plan trajectories to many candidates in parallel and only then score the grasp that each planned trajectory actually reaches. The claim is that the terminal configuration of a real planned motion, not the originally proposed target pose, is the object that should be evaluated for grasp success. The paper argues this removes a lose-lose tradeoff in the standard pipeline, where either the planner repeatedly reoptimizes toward lower-ranked grasps or relaxes its accuracy threshold and degrades the success estimate. In simulation the reordering improves success consistently across three generator architectures and two motion planners, and in real-world tests on eleven novel objects it raises success from 22% with the baseline to 80%, while also working on shelves and at different table heights.","feed_headline":"Plan first, score later: real grasp success hits 80%","feed_subtitle":"Evaluating the actual terminal pose, not the proposed target, lifts real-world success from 22% to 80% in new scenes.","key_machinery":"The mechanism that carries the argument is the vectorized motion planner, a planner that can solve many trajectory optimizations in parallel on a GPU; the paper uses Curobo and geometric fabrics for this role, and it is what makes planning to every proposed target cheap rather than sequential. The second component is the learned evaluator, a network trained on 28.9 million full-robot simulated grasp attempts, including hard negatives, that takes the object point cloud and a grasp and outputs a success likelihood; in FPTE the grasp fed to it is extracted from the last timestep of each planned trajectory, not from the generator's output. The paper deliberately drops the distance-to-target criterion, which is the load-bearing design choice: terminal poses that miss their targets are ranked and can be executed if the evaluator judges them likely to succeed, so planned trajectories are never discarded for failing to reach an arbitrary goal threshold.","core_discovery":"The paper's central discovery is that the traditional generate-evaluate-plan order is the source of a large, systematic grasp failure: generated targets are grounded in the object frame and often unreachable or colliding in the robot frame, so planning to the top-ranked target frequently fails, the pipeline replans to lower-ranked targets, and even when a relaxed planner reaches something close, it evaluates proximity to an idealized target rather than the quality of the actual end pose. FPTE, First Plan Then Evaluate, inverts this: a vectorized motion planner plans trajectories to all generated grasp targets simultaneously, and a learned evaluator scores the success likelihood of the terminal hand pose and configuration of each trajectory, ignoring distance to the target. The paper reports that this substitution yields higher predicted and actual success across all tested generator architectures (cVAE, diffusion, and MDN), both tested motion planners (Curobo and geometric fabrics), and 20 object shapes in simulation; in the real world, the predicted success of executed grasps averaged 90% versus 26% for the baseline, and actual success was 80% versus 22%, with the difference driven both by the baseline failing to find trajectories and by executing lower-quality grasps when it did.","pith_inferences":["The same plan-then-evaluate principle could apply to other robot manipulation skills where the planner's terminal state, not the commanded goal, determines outcome quality; for example placing, pushing, or tool use with loose goal tolerance.","Because the evaluator is the only judge of off-target terminal poses, the paper's reported real-world numbers are an upper bound on what the current evaluator supports; training the evaluator on varied table heights and clutter, rather than single isolated objects, would test whether the transferred success is limited by the evaluator's distribution.","An extension the paper does not run is FPTE with a purely geometric or analytical success score instead of the learned evaluator; if the ordering benefit persists, that would show the advantage comes from the plan-first order itself, not from the evaluator's learning.","The 90% average predicted likelihood for executed grasps suggests a practical intervention the paper mentions but does not deploy: a high-confidence execution gate with human handoff for low-confidence grasps could push system-level reliability well above the reported 80%."],"forward_implications":["Grasp success no longer depends on the generator proposing reachable grasps, since every proposed target is planned to and the terminal pose is what gets scored; this relaxes the burden on grasp generation.","The pipeline avoids sequential replanning through a ranked list; in the real-world baseline 23 of 55 attempts found no trajectory at all, whereas the parallel batch always provided an executable candidate.","The evaluator's calibration on terminal configurations becomes the central bottleneck, so any improvement in full-robot grasp-attempt data or evaluator training should translate directly into higher success.","Execution can be gated on the evaluator's confidence: the reported precision-recall analysis indicates that refusing grasps below a high predicted-likelihood threshold would raise success further.","The ordering benefit transfers across planner choices, generator architectures, and unseen environments without retraining the learned models, including novel objects, tables, shelves, and cluttered scenes."],"supporting_citations":[{"why":"supplies the heuristic grasp generation and the hand-closing and stiffness-raising execution routine used in data collection and deployment.","marker":"[8]"},{"why":"provides the cVAE generator architecture, the BPS point-cloud encoding, and the hard-negative strategy used to train the evaluator.","marker":"[21]"},{"why":"one of the diffusion-based grasp generators used in the cross-generator comparison of FPTE against the traditional pipeline.","marker":"[22]"},{"why":"another diffusion generator architecture used in the experiments and an example of the floating-end-effector training whose real-world gap FPTE addresses.","marker":"[23]"},{"why":"the vectorized motion planner that lets the pipeline plan to all grasp targets in parallel in the simulation experiments.","marker":"[30]"},{"why":"the geometric-fabrics vectorized planner used for parallel data collection and for the real-world FPTE experiments.","marker":"[31]"},{"why":"the MDN generator architecture in the cross-generator experiments, supporting the claim that the improvement holds across generator types.","marker":"[61]"}],"fun_headline_variants":["Plan to every grasp, then score the final pose: 80% real success","Invert the pipeline: plan first, evaluate later, real grasp success 80%","Multi-target planning beats generate-evaluate: 22% to 80% real success","Score terminal poses, not proposals: grasping hits 80% real world","Plan trajectories to all grasps, then pick best terminal pose: 80%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned success evaluator, trained only in simulation on single isolated objects on a fixed-height table, remains accurate when it scores real terminal configurations reached at different table heights, in shelves, under clutter, and at positions far from the originally proposed grasps.","fun_headline_variants_meta":{"raw":{"variants":["Plan to every grasp, then score the final pose: 80% real success","Invert the pipeline: plan first, evaluate later, real grasp success 80%","Multi-target planning beats generate-evaluate: 22% to 80% real success","Score terminal poses, not proposals: grasping hits 80% real world","Plan trajectories to all grasps, then pick best terminal pose: 80%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001015,"raw_usage":{"total_tokens":4344,"prompt_tokens":1060,"completion_tokens":3284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":3178}},"tokens_in":676,"tokens_out":3284,"duration_ms":18798,"temperature":1.0,"reasoning_tokens":3178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:12:16.034469+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FPTE in simulation with the goal tolerance deliberately widened so many planned trajectories terminate far from their targets, then compare evaluator scores against actual lift success separately for the off-target subset, without filtering out approach collisions; if the evaluator is no better than chance on that subset, the claimed advantage of terminal-configuration scoring collapses.","supporting_citations":[{"cited_title":"Planning visual-tactile precision grasps via complementary use of vision and touch,","cited_arxiv_id":null,"evidence_quote":"supplies the heuristic grasp generation and the hand-closing and stiffness-raising execution routine used in data collection and deployment."},{"cited_title":"Ffhnet: Generating multi-fingered robotic grasps for unknown objects in real-time,","cited_arxiv_id":null,"evidence_quote":"provides the cVAE generator architecture, the BPS point-cloud encoding, and the hard-negative strategy used to train the evaluator."},{"cited_title":"Dexdiffuser: Generating dexterous grasps with diffusion models,","cited_arxiv_id":null,"evidence_quote":"one of the diffusion-based grasp generators used in the cross-generator comparison of FPTE against the traditional pipeline."},{"cited_title":"Get a grip: Multi-finger grasp evaluation at scale enables robust sim-to-real transfer,","cited_arxiv_id":null,"evidence_quote":"another diffusion generator architecture used in the experiments and an example of the floating-end-effector training whose real-world gap FPTE addresses."},{"cited_title":"Geometric fabrics: Generalizing classical mechanics to capture the physics of behavior,","cited_arxiv_id":null,"evidence_quote":"the geometric-fabrics vectorized planner used for parallel data collection and for the real-world FPTE experiments."},{"cited_title":"Pick and place planning is better than pick planning then place planning,","cited_arxiv_id":null,"evidence_quote":"the MDN generator architecture in the cross-generator experiments, supporting the claim that the improvement holds across generator types."}],"review_version":2}