{"id":"49cf12c2-ec47-460d-b2e2-87a2aa9c87da","arxiv_id":"2605.23863","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A modified YOLO segmentation model plus sim-trained PPO control yields 84.3% overall success harvesting 281 strawberries in greenhouse trials on a real UR10e manipulator.","lead":"This paper builds a robotic strawberry harvester using an improved YOLO vision model for fruit detection in clutter and a PPO reinforcement learning policy trained in simulation to control a UR10e arm for reaching and picking. A generalist reader might examine it to see how sim-to-real transfer and task-specific perception can reduce the hardware trials needed for practical agricultural robots.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Isaac Lab sim fidelity to real fruit contact forces and greenhouse dynamics is the unverified prerequisite for the reported sim-to-real PPO transfer","rationale":"The reader's weakest assumption correctly isolates the sim-to-real gap as the single load-bearing condition for the DRL control claim. The greenhouse success rates only support the central engineering claim if that gap is small; the absence of any quantitative sim-real validation leaves the transfer unproven. No other internal inconsistency or unsupported derivation appears in the abstract or reported results.","tokens_in":1800,"tokens_out":375,"duration_ms":18439,"concrete_test":"Record 6-axis force-torque data and end-effector trajectories for 30 grasp-and-pull attempts on greenhouse strawberries using the deployed policy; replay the identical policy in Isaac Lab with the same initial poses and compare force magnitude histograms and peak detachment forces. If the real-world 95th-percentile force deviates >25% from the simulated distribution, the domain gap is large enough to question zero-shot transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (84.3% overall success on 281 strawberries) is produced by a PPO policy trained only in Isaac Lab and deployed zero-shot on the UR10e. For the 96.6% reaching and 91.3% grasp-and-pull rates to demonstrate successful sim-to-real transfer, the simulator must reproduce the relevant physics: variable stem detachment forces, fruit deformation under gripper contact, and lighting/occlusion statistics. The provided text gives no domain-randomization ranges, no force-torque or trajectory-matching metrics between sim and real, and no sensitivity analysis of the policy to parameter mismatch. Without these, the greenhouse numbers are consistent with successful transfer but do not rule out that performance depends on test conditions that happen to align with the simulation.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents a closed-loop robotic strawberry harvesting system that integrates a modified YOLO26-seg architecture (HRAttnEdge-YOLO26-seg) with high-resolution P2 branch, attention, and edge-supervised learning for instance segmentation in cluttered scenes, together with a target-conditioned PPO policy trained in Isaac Lab to generate joint-position commands for a UR10e arm. The policy is deployed zero-shot on physical hardware; greenhouse trials on 281 strawberries yield 96.6% reaching success, 91.3% grasp-and-pull success, and 84.3% overall harvesting success, with the vision model showing 10-14% segmentation gains on self-collected and public datasets and the controller producing smoother motion than an IK-based MoveIt baseline.","tokens_in":1970,"tokens_out":510,"duration_ms":17731,"significance":"If the sim-to-real transfer is substantiated, the work supplies concrete empirical evidence that a simulation-trained PPO policy can deliver high success rates on a physically deployed agricultural manipulator without real-world fine-tuning, together with a task-specific perception module that improves segmentation in occlusion-heavy scenes. The scale of the greenhouse evaluation (281 strawberries) and the direct baseline comparison constitute a tangible data point for sim-to-real manipulation in unstructured environments.","major_comments":[{"comment":"The central claim that the reported greenhouse performance demonstrates successful zero-shot sim-to-real transfer of the PPO policy rests on the unverified assumption that Isaac Lab reproduces the relevant contact forces, stem detachment dynamics, and lighting statistics. No domain-randomization parameter ranges, force-torque trajectory matching metrics, or sensitivity analysis of policy performance to simulation-reality mismatch are supplied, leaving the 84.3% overall success rate consistent with transfer but unable to rule out condition-specific alignment.","section":"Control and deployment description (greenhouse trials paragraph)"}],"minor_comments":[{"comment":"The abstract states success percentages without accompanying trial counts per metric, error bars, or statistical tests; these details appear only later in the greenhouse trials paragraph and should be summarized upfront for clarity.","section":"Abstract"},{"comment":"The claim of '10 to 14% improvement in segmentation performance' is stated without naming the exact metrics (mAP, IoU, etc.) or the precise baseline models against which the gain is measured.","section":"Perception results"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive feedback. We address the single major comment below and indicate the planned revision.","responses":[{"response":"We acknowledge that the original manuscript does not report the specific domain-randomization parameter ranges, force-torque matching metrics, or a sensitivity analysis. In the revised version we will add the exact randomization ranges used in Isaac Lab (lighting intensity and color temperature, stem stiffness and friction coefficients, fruit mass and size variation, and camera pose noise). We will also include a brief sensitivity study showing policy success rate versus selected randomization magnitudes. Direct force-torque trajectory matching is not feasible because the greenhouse trials did not instrument the UR10e with a force-torque sensor; we will therefore add an explicit limitations paragraph noting this gap and explaining that the 84.3 % success rate across 281 strawberries in an unstructured greenhouse, together with smoother motion than the MoveIt baseline, constitutes empirical support for transfer rather than conclusive proof against all possible sim-reality mismatches.","revision_made":"yes","referee_comment":"The central claim that the reported greenhouse performance demonstrates successful zero-shot sim-to-real transfer of the PPO policy rests on the unverified assumption that Isaac Lab reproduces the relevant contact forces, stem detachment dynamics, and lighting statistics. No domain-randomization parameter ranges, force-torque trajectory matching metrics, or sensitivity analysis of policy performance to simulation-reality mismatch are supplied, leaving the 84.3% overall success rate consistent with transfer but unable to rule out condition-specific alignment."}],"tokens_in":1464,"tokens_out":330,"duration_ms":20014,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper gives greenhouse numbers for a strawberry harvester that uses a modified YOLO for vision and PPO trained in simulation for the arm control, hitting 84.3% overall success across 281 fruits. The system combines HRAttnEdge-YOLO26-seg with target-conditioned PPO on a UR10e and reports the reaching and grasping sub-success rates separately. It does well by providing a complete pipeline with reported improvements in segmentation accuracy of 10 to 14 percent over other methods on both custom and public datasets. The controller tests show more stable motion than the MoveIt inverse kinematics baseline, and the greenhouse trial supplies concrete success percentages that allow direct comparison to other harvesting approaches. The soft spots center on the sim-to-real claim. The abstract states the policy was trained only in Isaac Lab and deployed without fine-tuning, but it gives no domain randomization ranges, no force or trajectory matching metrics between sim and real, and no sensitivity analysis. Without that, the high success rates could partly reflect test conditions that match the simulation rather than robust transfer. The missing error bars and trial count details also make the 84.3 percent figure harder to interpret for repeatability. This kind of paper is aimed at robotics researchers and engineers focused on agricultural automation who need practical examples of combining perception and learning-based control. It is worth sending for peer review. The physical results provide enough substance for referees to assess the contribution and request additional validation on the simulation aspects if needed.","headline":"Greenhouse results show workable sim-to-real control for strawberry picking, but the transfer relies on an unverified simulation match.","tokens_in":2489,"tokens_out":367,"would_cite":false,"duration_ms":30181,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Robotic harvesting pipeline with PPO sim-to-real has no structural overlap with recognition-cost or distinction-forcing machinery","alignment":"orthogonal","rationale":"The paper's central machinery (HRAttnEdge-YOLO26-seg segmentation + target-conditioned PPO policy trained in Isaac Lab for UR10e reaching) is standard empirical robotics. It neither invokes nor parallels any RS element such as J-cost, phi-ladder, 8-tick periodicity, or the reality_from_one_distinction theorem. No passages reference recognition logic, reciprocal cost, or parameter-free constant derivations.","tokens_in":56692,"confidence":"high","tokens_out":137,"duration_ms":5912,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A robot arm harvests strawberries at 84.3 percent overall success after its control policy is trained only in simulation.","keywords":["strawberry harvesting","robotic manipulation","deep reinforcement learning","sim-to-real transfer","instance segmentation","PPO control","greenhouse robotics","UR10e manipulator"],"falsifier":"Deploy the same PPO policy on the UR10e in the greenhouse and record whether the combined reaching-plus-grasp success rate falls materially below the reported 91.3 percent because of unmodeled contact forces or lighting changes.","tokens_in":2705,"feed_emoji":"🍓","tokens_out":722,"duration_ms":15957,"temperature":0.7,"pith_summary":"The paper builds a complete harvesting pipeline that pairs a custom vision model for locating fruit in clutter with a reinforcement learning controller that learns reaching and grasping motions entirely inside a simulator. The controller then runs directly on a physical UR10e arm inside real greenhouses, producing joint commands without any real-world policy updates or traditional motion planners. Across 281 strawberries the system reached 96.6 percent success on the reach phase, 91.3 percent on grasp-and-pull, and 84.3 percent end-to-end. A reader should care because the approach replaces costly physical trial-and-error with cheaper simulation training while still delivering usable performance in unstructured agricultural scenes.","feed_headline":"Simulation-trained robot harvests strawberries at 84.3% success","feed_subtitle":"PPO policy learned only in Isaac Lab plus improved vision delivers closed-loop picking on a UR10e without real-world fine-tuning.","key_machinery":"The target-conditioned PPO policy trained in Isaac Lab that maps fruit location observations to smooth joint-position commands for the UR10e arm.","core_discovery":"The paper shows that a target-conditioned PPO policy trained in Isaac Lab to output joint-position commands, when combined with the HRAttnEdge-YOLO26-seg perception model, produces stable closed-loop harvesting on a UR10e manipulator that reaches 84.3 percent overall success on 281 strawberries in greenhouse conditions, outperforming an inverse-kinematics MoveIt baseline in motion smoothness and eliminating the need for exhaustive real-robot data collection before deployment.","pith_inferences":["The same simulation-trained policy structure could be reused for other soft or clustered fruits if the contact-force model in simulation is adjusted accordingly.","Adding camera calibration drift detection at runtime would test whether the reported success rates remain stable across multiple days of greenhouse operation.","The approach implies that scaling the number of simulated environments could reduce the remaining 15.7 percent failure rate without collecting new real-robot failures."],"forward_implications":["The vision model raises segmentation accuracy by 10 to 14 percent over baseline YOLO variants on both in-house and public datasets.","The PPO controller produces smoother joint trajectories than the MoveIt IK baseline in controlled lab tests.","The full pipeline reduces hardware dependency by training the controller exclusively in simulation before direct real-robot deployment.","The integrated system achieves 84.3 percent end-to-end success on 281 strawberries without planner-dependent reaching."],"fun_headline_variants":["Sim-trained PPO harvests strawberries at 84.3% success","Isaac Lab trained PPO achieves 84.3% strawberry harvest success","Target-conditioned PPO reaches 84.3% in real strawberry picking","Sim-to-real control yields 84.3% robotic strawberry harvesting"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The Isaac Lab simulator reproduces real greenhouse lighting, fruit dynamics, and robot contact forces closely enough that a policy trained only inside it transfers to the physical robot without extra fine-tuning or new failure modes.","fun_headline_variants_meta":{"raw":{"variants":["Sim-trained PPO harvests strawberries at 84.3% success","Isaac Lab trained PPO achieves 84.3% strawberry harvest success","Target-conditioned PPO reaches 84.3% in real strawberry picking","Sim-to-real control yields 84.3% robotic strawberry harvesting"]},"model":"grok-4.3","cost_usd":0.006617,"raw_usage":{"total_tokens":3130,"prompt_tokens":751,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":66174500,"prompt_tokens_details":{"text_tokens":751,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2307,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":751,"tokens_out":72,"duration_ms":24448,"temperature":1.0,"reasoning_tokens":2307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-25T03:46:02.910110+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Deploy the same PPO policy on the UR10e in the greenhouse and record whether the combined reaching-plus-grasp success rate falls materially below the reported 91.3 percent because of unmodeled contact forces or lighting changes.","supporting_citations":[],"review_version":1}