{"id":"b55bef44-3d80-4843-ab00-70cda8699864","arxiv_id":"2506.23369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A viewpoint-planning algorithm constrains the camera search to a 1D picking ring and achieves 100 percent simulated success in two avocado harvesting scenarios.","lead":"This paper presents GS-NBV, an algorithm that plans camera viewpoints for robotic avocado harvesting by restricting the search to a 1D ring around the fruit and scoring views by how much of the fruit and stem are visible. In simulation it reports 100 percent success in two occlusion-heavy case studies, outperforming two comparison planners.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 100% success rate is measured against the planner's own picking score, which excludes the axis error that the paper itself reports at 22–30°; this makes the harvesting-relevant claim unsupported.","rationale":"The reader's conditional verdict already captures the main risk: the picking score is not validated as a proxy for real avocado harvesting success. My stress-test sharpens this into a specific, internally visible problem: the score omits fruit-axis error, and the paper's own Table III reports axis errors at or near the 30° bounding cap. Because the picking ring and the viewpoint sampling space are constructed from the estimated axis, a 30° axis error means the planned viewpoints are not necessarily aligned with the ground-truth picking geometry. The paper's statement that this 'does not hinder their ability to identify viable picking views' is the exact unsupported bridge between the reported 100% success rate and the harvesting-relevant claim. The proposed test would settle whether the success rate survives an independent, ground-truth-based definition of pickability. I do not see a separate internal inconsistency in the planning algorithm itself; the algorithm appears to do what it is formally asked to do. The concern is about whether the formal objective is the right objective, which is why the verdict should remain conditional rather than be upgraded or rejected outright.","tokens_in":9372,"tokens_out":5756,"duration_ms":63695,"concrete_test":"Using the released code and the same two Gazebo scenarios, replace the success criterion with a ground-truth harvesting-relevant condition: final success requires (i) estimated axis error below a threshold such as 10°, and (ii) the final camera pose is within the canonical/front view cone defined by the true fruit axis and has a clear line of sight to the peduncle attachment point. If GS-NBV's success rate drops materially below 100% (or iterations exceed two), the reported success is an artifact of optimizing its own proxy rather than finding a truly pickable view.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GS-NBV consistently finds a 'pickable view' is supported only by the internal metric spick = sdis * (1 - socc) with threshold 0.9 (Section II, Algorithm 1). This metric checks visibility and occlusion of the fruit contour, but contains no term for 6D pose accuracy, grasping configuration, or peduncle reachability. Table III reports fruit axis error of 22.63 ± 7.37° in Group 1 and exactly 30.00 ± 0.00° in Group 2; the latter is the artificial cap introduced in Section III-A.2, so in Group 2 the axis estimate carries no information. The paper itself concedes this 'poses additional challenges for the final picking step' but asserts it 'does not hinder their ability to identify viable picking views.' That assertion is the load-bearing step: a view can score above 0.9 while the estimated picking ring (normal = estimated axis) is tilted by 30° from the true vertical axis, so the final view need not be a canonical/front view usable by the end-effector. Indeed, Section IV-D shows GS-NBV's final Group-2 view is 'from below the fruit,' which is not the canonical/front view the Introduction argues is required. Thus the 100% success rate validates that the planner reaches its self-defined score, not that it identifies an avocado-harvesting viewpoint.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GS-NBV, a geometry-based, semantics-aware viewpoint planner for avocado harvesting in occluded settings. The planner detects a partially visible avocado, estimates its position and axis, constrains candidate camera viewpoints to a 1D picking ring around the fruit, samples four viewpoints on that ring, and scores them with a utility function that combines semantic information gain, a visibility-and-occlusion picking score spick, and a distance term. The camera is moved iteratively until spick > 0.9. The method is evaluated in a Gazebo simulation against SC-NBVP and GNBV in two static groups (one with an added occlusion board), with 10 trials per group. The paper reports a 100% success rate for GS-NBV, lower planning iterations, and faster sampling-based planning than SC-NBVP, alongside open-source code.","tokens_in":9740,"tokens_out":5353,"duration_ms":57510,"significance":"The algorithmic idea of reducing the viewpoint search space to a 1D ring and evaluating candidate views with a cheap picking score is sensible, and the public code is a concrete strength. The efficiency comparison and the qualitative demonstration that the planner can escape occluded initial views are plausible. However, the central harvesting claim is not established: the success metric is identical to the planner's own stopping condition and contains no term for pose accuracy, grasping configuration, or peduncle reachability, while the reported fruit-axis errors are 22-30 degrees and the final Group-2 view is from below the fruit. The contribution is therefore best described as a viewpoint planner that reliably reaches its internal pickability condition, with the step from that condition to actual avocado harvesting remaining unsupported.","major_comments":[{"comment":"The evaluation's 'success rate' is self-referential: a run is counted as successful exactly when spick = sdis * (1 - socc) exceeds 0.9, which is the same condition under which the while loop in Algorithm 1 terminates. Consequently, the reported 100% success rate in Section IV-D measures how often the planner reaches its own internal stopping condition, not whether the resulting viewpoint enables avocado harvesting. The abstract and conclusion present this rate as harvesting success; supporting that claim requires an external metric, such as an actual simulated or physical picking attempt, or at least a demonstration that spick correlates with end-effector picking success.","section":"Section II, Algorithm 1, Section IV-D"},{"comment":"The picking score spick contains no term for 6D pose accuracy, grasping configuration, or peduncle reachability, yet Section IV-E asserts that the large fruit-axis error 'does not hinder their ability to identify viable picking views.' Table III reports axis errors of 22.63 +/- 7.37 degrees in Group 1 and exactly 30.00 +/- 0.00 degrees in Group 2, with the latter value equal to the artificial cap introduced in Section III-A.2. With such errors, the estimated picking ring is tilted well away from the true ring, and Section IV-D reports that GS-NBV's final Group-2 view is 'from below the fruit,' which is not the canonical/front view that Section I motivates as necessary for avocado picking. The load-bearing assertion that spick > 0.9 implies a harvestable view is therefore unsupported; the authors should either incorporate axis error into the success metric or validate the proxy against a concrete picking action.","section":"Section II, Section III-A.2, Table III, Section IV-E"},{"comment":"The evidence base is too thin for the robustness claims made in the abstract and conclusion. Only two static scenes with a single fruit were tested, each with 10 trials and no statistical tests or confidence intervals on success rates. The baselines were fine-tuned by the authors, and the comparison does not control for prior geometric information: GS-NBV is given the picking-ring radius and a fruit-centered ROI, while the baselines search larger or differently structured spaces. This may explain both the speed advantage in Section IV-C and the success-rate margin, but it makes the reported 100% versus 30-60% improvement difficult to interpret. Additional trials, multiple fruit poses and occlusion configurations, and ablations that remove or perturb the ring prior are needed to substantiate the claimed advantage.","section":"Section IV-A, Section IV-C, Table I"}],"minor_comments":[{"comment":"The text calls lambda a 'motion cost penalty' and writes the utility as Gsem * e^{-lambda * d}, but Table I sets lambda = -1, making the factor e^{d} and thus rewarding more distant viewpoints. Please either rename the coefficient or clarify the sign convention.","section":"Section III-C.1, Table I"},{"comment":"The YOLOv8 instance segmentation is fine-tuned on 108 images with leave-one-out cross-validation, but no detection or segmentation accuracy (e.g., mAP, F1, or mask IoU) is reported; please add these metrics to support the perception pipeline's reliability.","section":"Section III-A.1"},{"comment":"The sentence 'the time consumption of GNBV is zero' is ambiguous: Table II shows zero only for the viewpoint-sampling subprocess, while the overall NBV planning time for GNBV is 0.461 seconds. Please rephrase to indicate that only the sampling time is zero.","section":"Section IV-C"},{"comment":"The number of planning iterations is reported as a single mean without variance, and the success-rate explanation for SC-NBVP's improvement in Group 2 would benefit from clarification: the text says right-side viewpoints 'generate high SC scores' and are 'less likely to be selected,' which seems contradictory; please spell out the causal mechanism.","section":"Table III, Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies on the authors' own prior end-effector work [12] to define the picking ring and canonical/front views, but the current experiments never exercise the end-effector, so the harvesting claim cannot be independently validated here. The algorithmic core is well defined and the code is public, so I would not recommend rejection; however, the evaluation section should be restructured around an external success metric rather than the planner's internal stopping condition, and the abstract/conclusion should be tempered accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this paper proposes a next-best-view planner for avocado harvesting that restricts viewpoint sampling to a 1D 'picking ring' around the fruit. That's a sensible geometric reduction and the combination is new. The authors integrate it with a semantic OctoMap and a scalar picking score that combines discoverability and occlusion. The code is public, and the simulation results are clearly reported: 100% success in two static scenes, with about 2.5x speedup over one sampling baseline. The time breakdown is useful.\n\nThe soft spots are real. The success metric is the planner's own picking score, which checks whether the fruit and peduncle are visible and the surrounding curve is mostly unoccluded. It has no term for the estimated fruit axis, and the paper reports axis errors of 22–30°. In Group 2 the axis error is exactly 30.00±0.00°, which is the artificial cap the authors impose. They acknowledge the axis error 'poses additional challenges for the final picking step' but claim it does not hinder identifying viable views. That claim is not supported by the reported data—in fact, the Group-2 final view for GS-NBV is described as 'from below the fruit,' which is not the canonical view that the introduction says is required. So the 100% success rate validates that the planner reaches its own threshold, not that it produces a harvestable viewpoint. The evidence base is thin: two scenes, ten trials each, no physical tests, and the baselines were fine-tuned by the authors, with one baseline's search space reduced to 2D.\n\nTo the paper's credit, the limitations are not hidden—they are stated in the results and future work. The algorithm is described well enough to reimplement, and the 1D sampling idea is a genuine contribution that could transfer to other fruits with a well-defined picking geometry. The axis error issue is known to the authors and is the obvious reason this needs physical validation.\n\nMy take: this is a solid workshop/IROS-quality paper with a clear but unvalidated core claim. It deserves a serious referee, because the approach is worth engaging with and the shortfall is in validation, not in the idea. The next step is either physical picking trials or a success metric that includes pose error and reachability. I would not cite it as evidence for harvesting success, but I would cite it for the 1D picking-ring viewpoint-sampling idea. Bring it to the reading group if you want a case study in circular evaluation metrics.","headline":"A clever 1D viewpoint-sampling idea for avocado picking, but the 100% success claim only validates the planner's own picking score, not actual harvestability.","tokens_in":10222,"tokens_out":4495,"would_cite":true,"duration_ms":43701,"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 constraining viewpoint search to a 1D picking ring and scoring views by fruit-stem visibility and occlusion lets a robot arm find a pickable avocado view under heavy occlusion within two planning iterations.","keywords":["viewpoint planning","next-best view","avocado harvesting","occlusion handling","picking score","semantic octomap","1D picking ring"],"falsifier":"Run the same planner in a simulated or physical orchard where the avocado hangs with its true axis more than 30 degrees from vertical, or attempt an actual grasp after the planner reports a pickable view; if the reported 100 percent pickable-view success does not lead to successful peduncle cuts because the axis error pushes the gripper into the fruit, the central claim that the picking score identifies a harvestable viewpoint would be falsified.","tokens_in":9183,"feed_emoji":"🥑","tokens_out":7976,"duration_ms":69603,"temperature":0.7,"pith_summary":"The paper tries to establish that the hardest part of robotic avocado harvesting under occlusion—finding a camera pose from which the fruit and its peduncle are clearly visible—can be solved by geometry rather than brute-force search. It proposes GS-NBV, a next-best-view planner that shrinks the search to a one-dimensional picking ring around the estimated fruit position, samples four viewpoints on that ring, and scores them with a visibility-and-occlusion metric. The claim is that this small, structured search reliably reaches a view that makes the fruit pickable, defined as a picking score above 0.9, in an average of 1.2 to 1.5 planning iterations. A sympathetic reader would care because avocados cannot be grabbed anywhere on the surface: the gripper needs a specific canonical or front view of the fruit and peduncle, and occlusions in unstructured trees make that view hard to find.","feed_headline":"Planner finds a clear avocado view in two planning steps","feed_subtitle":"A geometry-based planner shrinks the camera search to a 1D ring and achieves 100% success in simulated occlusion tests.","key_machinery":"The load-bearing object is the picking ring: the 1D circle in the fruit's $x$-$y$ plane that defines where an avocado may be grasped, and the same circle is used as the viewpoint search space. The paper adds a picking score $s_{\\mathrm{pick}} = s_{\\mathrm{dis}}(1 - s_{\\mathrm{occ}})$, where $s_{\\mathrm{dis}}$ is a binary discoverability flag requiring the fruit and peduncle to be visible and overlapping or within five pixels, and $s_{\\mathrm{occ}}$ is the fraction of the fruit's surrounding contour curve that has depth closer than the fruit's minimum depth plus 0.06 m. A fruit is declared pickable when $s_{\\mathrm{pick}} > 0.9$. The planner's utility score for ranking viewpoints is the expected semantic information gain from the semantic OctoMap multiplied by $e^{-\\lambda d}$, with $\\lambda = -1$ so that more distant viewpoints are preferred; this promotes exploration along the ring while occlusion measurements shrink the candidate arc from both sides. This combination reduces raycasting cost and gives the planner its two-iteration convergence.","core_discovery":"On the paper's own terms, the central discovery is that the avocado's picking geometry—a circumferential ring around the fruit where a harvesting end-effector can grasp and detach it—also organizes viewpoint search. Starting from a partially occluded view, the planner detects the fruit, estimates its position and axis, builds a semantic occupancy map of the region of interest, and restricts candidate camera poses to a 0.21 m radius circle around the fruit, minus the arc behind the tree and minus arcs already blocked by newly discovered obstacles. Four evenly spaced viewpoints on that shrinking arc are evaluated by a utility score that combines expected semantic information gain and motion cost, and the camera moves to the best one. The paper reports that this procedure achieved a 100% success rate in both of its simulated occlusion scenarios, with a pickable view found within two planning iterations, whereas the two comparison methods succeeded in at most 60 percent of trials.","pith_inferences":["The authors leave implicit that the 1D-ring reduction should transfer to other peduncle-cut fruits, such as litchi, grapes, or peppers, whenever the graspable region forms a ring around the fruit.","A physical harvest would still face the 22-30 degree fruit-axis error, so the reported 100 percent success is about finding a clear viewpoint, not about a successful grasp; real picking would likely require re-orienting or extra observations before the end-effector contacts the fruit.","A direct test of the geometric prior would be to hang the avocado with its axis more than 30 degrees from vertical; the planner's bounded-axis assumption predicts failure or a misleading ring, which would expose the method's limit."],"forward_implications":["In both simulated occlusion scenarios, the planner reports a picking score above 0.9 in every trial, needing on average 1.5 planning iterations in the first group and 1.2 in the second.","Restricting the search to a 1D ring makes each viewpoint-evaluation pass take about 1.5 s, roughly 2.5 times faster than the sampling-based comparison method.","A fruit position estimate with mean error of 0.02-0.04 m is sufficient to define a reliable picking ring, even though the fruit axis estimates carry 22-30 degree errors.","The planner's output is directly usable by an end-effector designed for canonical and front avocado views, since the picking score only accepts views in which the fruit and peduncle appear together."],"supporting_citations":[{"why":"Provides the instance-segmentation model that produces masks of the avocado and peduncle for pose estimation.","marker":"[25]"},{"why":"Defines the canonical and front views and the end-effector design that make the picking ring the right search structure.","marker":"[12]"},{"why":"Supplies the semantic OctoMap representation and the gradient-based local next-best-view approach whose utility score this work extends.","marker":"[26]"},{"why":"Provides the raycasting operation used to update the occupancy map and test viewpoint visibility.","marker":"[27]"},{"why":"Provides the max-fusion rule that merges new semantic observations into the OctoMap.","marker":"[28]"},{"why":"Supplies the sampling-based comparison method whose random viewpoint search is reduced from 3D to 2D in the experiments.","marker":"[29]"},{"why":"Supplies the viewpoint dissimilarity filtering that discards a newly sampled viewpoint within 0.1 m of a previously evaluated one.","marker":"[20]"},{"why":"Provides the motion-planning layer that moves the arm from the current viewpoint to the next one.","marker":"[31]"}],"fun_headline_variants":["1D ring confines camera search for occluded avocado picking","Avocado viewpoint planner: geometry ring, 100% simulation success","Shrinking viewpoint search to a ring boosts avocado harvesting","Two-step planner finds avocados behind occlusions at 100% rate","Geometry and semantics combine for perfect avocado views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the picking score—fruit and peduncle visible and the surrounding contour curve at least 90 percent unoccluded—is a sufficient stand-in for a real harvestable grasp, even though the estimated fruit axis can be off by 22 to 30 degrees.","fun_headline_variants_meta":{"raw":{"variants":["1D ring confines camera search for occluded avocado picking","Avocado viewpoint planner: geometry ring, 100% simulation success","Shrinking viewpoint search to a ring boosts avocado harvesting","Two-step planner finds avocados behind occlusions at 100% rate","Geometry and semantics combine for perfect avocado views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3275,"prompt_tokens":907,"completion_tokens":2368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2285}},"tokens_in":523,"tokens_out":2368,"duration_ms":18549,"temperature":1.0,"reasoning_tokens":2285,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:44:06.824669+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same planner in a simulated or physical orchard where the avocado hangs with its true axis more than 30 degrees from vertical, or attempt an actual grasp after the planner reports a pickable view; if the reported 100 percent pickable-view success does not lead to successful peduncle cuts because the axis error pushes the gripper into the fruit, the central claim that the picking score identifies a harvestable viewpoint would be falsified.","supporting_citations":[{"cited_title":"Design of an end-effector with application to avocado harvesting,","cited_arxiv_id":null,"evidence_quote":"Defines the canonical and front views and the end-effector design that make the picking ring the right search structure."},{"cited_title":"Gradient-based local next-best-view planning for improved perception of targeted plant nodes,","cited_arxiv_id":null,"evidence_quote":"Supplies the semantic OctoMap representation and the gradient-based local next-best-view approach whose utility score this work extends."},{"cited_title":"Real-time voxel based 3d semantic mapping with a hand held rgb-d camera,","cited_arxiv_id":null,"evidence_quote":"Provides the max-fusion rule that merges new semantic observations into the OctoMap."},{"cited_title":"View planning for grape harvesting based on active vision strategy under occlusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the sampling-based comparison method whose random viewpoint search is reduced from 3D to 2D in the experiments."},{"cited_title":"Nbv-sc: Next best view planning based on shape completion for fruit mapping and reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the viewpoint dissimilarity filtering that discards a newly sampled viewpoint within 0.1 m of a previously evaluated one."},{"cited_title":"[Online]","cited_arxiv_id":null,"evidence_quote":"Provides the motion-planning layer that moves the arm from the current viewpoint to the next one."}],"review_version":1}