{"id":"e688cad8-1d42-42ab-96d8-4e23038f52b3","arxiv_id":"2507.23015","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A simulation-trained optical-flow controller guides a robot cutter to pruning points and orients it perpendicular to branches, with about 30% success in simulation and 35% on a lab tree.","lead":"This paper trains a robot arm to prune apple tree branches using optical flow from a wrist-mounted camera, rather than a full 3D model. It matters because pruning is a major orchard labor cost, and the results show a vision-based policy can transfer from simulation to a physical tree, though success rates remain low.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-shot transfer claim is only validated on a lab proxy tree with artificial connectors; an actual orchard deployment is untested, so the 30-35% success rate may not represent real V-Trellis Envy conditions.","rationale":"The reader's weakest assumption is that the synthetic tree generator is realistic enough for zero-shot transfer. My concern is related but distinct: the real-world validation itself is performed on a simplified lab proxy, so even a perfectly realistic simulator would not establish the stated real-world claims. The reader's rationale does mention the lab proxy as an unverified assumption, so there is partial overlap, but I elevate it to the primary load-bearing issue because it directly undermines the abstract's 'real-world' deployment claim and the comparative result against point-cloud planning. The paper's contribution is still valuable as a sim-to-lab-transfer study, and the simulation experiments are substantial, so conditional acceptance remains appropriate. The concrete test is an actual orchard deployment with a fair planner baseline, which would resolve the concern.","tokens_in":9707,"tokens_out":7295,"duration_ms":87158,"concrete_test":"Run the trained policy on an actual V-Trellis Envy tree in an outdoor orchard, using the same UR5/cutter/camera hardware and no manual adjustments beyond the cutpoint coordinates, and evaluate at least 50 cutpoints selected by a horticultural expert, reporting success rate with a 95% confidence interval. For the baseline, reconstruct the full tree with a depth camera under the same lighting, automatically compute goal configurations, and give RRT-Connect the same 60-second budget. If the policy's success rate is significantly below the 30-35% range, or if RRT-Connect achieves comparable or higher success, the paper's central claims of zero-shot transfer and real-world superiority over point-cloud planning are not confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a policy trained in simulation transfers zero-shot to a real V-Trellis Envy tree and outperforms point-cloud-based RRT-Connect in the real world. However, the real-world evaluation is conducted on a proxy tree built in the lab from real apple branches attached with custom plastic connectors that control branch location and orientation (Section V-B, Figure 5). This eliminates the natural variation in branch attachment, clutter density, and tie-down geometry that the simulator was designed to reproduce. The environment also includes a printed cutter, a backdrop image of an orchard, and no outdoor lighting or weather. With only 20 trials, the 35% success rate (7/20) has a wide confidence interval (roughly 15-59%) and is not statistically distinguishable from the 30% simulation result. The RRT-Connect comparison is further weakened: the point cloud covers only one branch, goal configurations are manually supplied 5 cm from the cutpoint, and the planner is given only 60 seconds. Thus the claim that the method outperforms point-cloud planning in the real world is not supported by a controlled comparison. If an actual orchard has denser clutter, natural branch angles, or different optical-flow response to sunlight, the policy's success rate could be substantially lower. The simulator-realism assumption is important, but even a perfect simulator cannot rescue an evaluation that does not test the claimed deployment scenario.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a closed-loop visuomotor controller for robotic dormant pruning in V-trellis apple orchards. The controller consumes optical flow images from a wrist-mounted camera, a goal specification for the cut point, and robot proprioception, and outputs 6DOF end-effector velocities. Training is done with PPO in a PyBullet simulator whose trees are generated by an L-Py extension that models annual pruning and tie-downs. The reward function combines reaching, pointing, perpendicularity, collision, and slack terms. In simulation, the policy achieves 30.06% success over 3000 trials against an oracle RRT-Connect baseline with perfect geometry at 60.4%. In the real world, the policy attains 35% success over 20 trials on a laboratory proxy tree built from real apple branches with custom connectors. A point-cloud-based RRT-Connect baseline succeeds on 2 of 10 trials on a single reconstructed branch. The paper also provides a task-space analysis of failure modes as a function of branch orientation. The central claim is that optical-flow-based visuomotor control transfers zero-shot from simulation to a real tree and outperforms point-cloud-based planning in the real world.","tokens_in":10037,"tokens_out":4081,"duration_ms":48132,"significance":"If the central claim were fully supported, the paper would make a useful contribution: it offers a perception approach that avoids fragile 3D reconstruction of thin branches, an openly described synthetic tree generator that addresses a gap in existing L-systems tools, and a sensible oracle-planner upper bound for evaluating learned policies. The failure-mode analysis in Section VI is a step toward understanding where learned pruning policies fail. However, the real-world evidence is preliminary and the headline claims exceed what the experimental design can support, because the real-world environment is a lab proxy with artificial connectors and the baseline planner is not run under comparable conditions.","major_comments":[{"comment":"The real-world evaluation is conducted on a proxy tree built in the laboratory from real apple branches attached with custom plastic connectors that control branch location and orientation, with an orchard backdrop and a printed cutter. This removes the natural variability in branch attachment, clutter density, tie-down geometry, lighting, and weather that the simulator was designed to model. The abstract's statement that the policy is deployed 'in the real-world for an example V-Trellis envy tree' is therefore an overstatement. Please either evaluate on an actual orchard tree or explicitly reframe all real-world claims as validation on a lab proxy, not as orchard deployment.","section":"Section V-B, Figure 5"},{"comment":"Only 20 real-world trials are reported, with 7 successes (35%). The exact binomial 95% confidence interval for this result is approximately 15% to 59%, which is so wide that the result is not statistically distinguishable from the 30% simulation success rate. The claim of 'matching the results in simulation' and 'high sim-to-real transferability' is not supported by this sample size. Please report confidence intervals or credible intervals for all success rates and avoid claims of equivalence based on point estimates.","section":"Section V-B, Real-World Results"},{"comment":"The RRT-Connect baseline is not a controlled comparison. It reconstructs only one branch rather than the full tree, uses manually supplied collision-free goal poses 5 cm from the cutpoint, is limited to 60 seconds of planning time, and is executed only for cutpoints that are visible in the point cloud. With 10 trials and 2 successes, the statement 'our visuomotor policy outperforms point cloud based planners' is not supported. A fair comparison would give both methods the same cutpoints, the same perception input, and the same success criteria, or the claim should be limited to the observation that the point cloud failed to capture some tertiary branches.","section":"Section V-B, RRT-Connect comparison"},{"comment":"The authors state that the synthetic tree geometry 'has not been formally verified [to be] statistically consistent with real orchards.' This is a load-bearing limitation because the paper's central claim is zero-shot transfer from simulation to the real world. Without quantitative evidence that the simulated branch diameter, length, orientation, curvature, and clutter distributions match commercial V-Trellis Envy orchards, the reported 30-35% success rates may not transfer beyond the specific lab proxy tested. Please provide a statistical comparison between the synthetic tree bank and orchard scans (e.g., branch diameter/length histograms, orientation distributions, clutter density), or temper the zero-shot transfer claim accordingly.","section":"Section III, 'Simulating the Reaching Task'"}],"minor_comments":[{"comment":"The cultivar name should be capitalized: 'V-Trellis Envy tree', not 'V-Trellis envy tree'.","section":"Abstract"},{"comment":"For the real-world trials, only 20 points are shown for each error distribution; consider adding a bootstrapped confidence band or a violin-style plot so the reader can judge the spread without over-interpreting the density curves.","section":"Figure 4"},{"comment":"The perpendicularity reward uses the absolute value of the cosine similarity, which rewards the cutter's left-right vector being aligned with either direction of the branch. If the cutter jaws are symmetric so that anti-parallel alignment is also acceptable, please state this explicitly; otherwise clarify the intended direction.","section":"Section IV-C, Equation (3)"},{"comment":"The reward weights alpha_m = 5, alpha_p1 = 6, and alpha_p2 = 2 are described only as 'empirically determined.' A sentence on the selection procedure or a small sensitivity study would improve reproducibility.","section":"Section IV-C"},{"comment":"The oracle RRT-Connect success rate is reported as 60.4% over 3000 cutpoints, but the description of goal-configuration sampling could be clearer: state explicitly that the 60.4% is the fraction of cutpoints for which at least one of the 100 sampled goal configurations has a collision-free path, and whether the same 3D locations are used as for the RL policy.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The simulation study and the oracle-planner upper bound are the paper's strongest parts, and the synthetic tree generator is a genuine contribution. The real-world validation, however, is the weakest link: a 20-trial lab proxy with custom connectors cannot support the abstract's 'real-world deployment' claim, and the RRT-Connect comparison is not apples-to-apples. I would be willing to accept a revised version that either presents real orchard results or substantially tempers the claims and adds confidence intervals and simulator-realism validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: this is a real step forward for optical-flow visuomotor pruning, but the zero-shot transfer claim is braver than the evidence. The paper itself is transparent about the lab proxy and the unverified simulator realism, so the gap is in the abstract and conclusion, not in the body.\n\nWhat's actually new: this is the first extension I know of that takes the You et al. optical-flow pruning idea from 2D perpendicular approach to full 6DOF with visible obstacle avoidance. The modified L-Py generator with cyclic pruning and tie-downs is a solid piece of engineering, and the simulation evaluation is decent: 3000 trials, an oracle RRT-Connect upper-bound comparison, and a task-space failure analysis that correctly identifies the UR5's reachability limits and self-collision zones. The reward design is careful, and using RAFT optical flow for sim-to-real is consistent with their prior work.\n\nSoft spots, in rough proportion: first, the real-world evaluation is on a lab tree with custom plastic connectors that fix branch location and orientation. That removes exactly the variation the simulator was trying to capture. The paper says so in Section V-B, but the abstract still says 'zero-shot transfer' and 'deploy in the real world.' With 20 trials, 35% is not distinguishable from the 30% simulation number, and there are no confidence intervals. Second, the RRT-Connect comparison does not support 'outperforms point-cloud planners.' The point cloud covers one branch, goal poses are manually supplied 5cm from the cutpoint, and the planner gets 60 seconds; 5 of 10 cutpoints were missing from the cloud entirely. The comparison is suggestive, not controlled. Third, Section III concedes the simulated tree geometry is not formally verified against real orchards. That is the right kind of honesty, but it is a load-bearing assumption for the whole sim-to-real story. Minor: no code or data release, and the real-world runs are 4x slower than sim.\n\nWho is this for? People building orchard simulators or working on optical-flow control for thin structures will get real value. The failure analysis and the L-Py extension are the strongest parts. The deployment claim should be read as a lab feasibility study, not an orchard demo.\n\nMy recommendation: send it to peer review — it's a serious, well-written paper that deserves referee time. But I'd push for major revision: temper the zero-shot language, add confidence intervals, release the code/models, or better, run a small pilot in an actual V-Trellis orchard. Desk rejection would be a mistake.","headline":"Solid simulator and 6DOF optical-flow controller, but the zero-shot transfer claim is only supported by a lab proxy, not an orchard, and the real-world planner comparison is too weak for the 'outperforms' claim.","tokens_in":10507,"tokens_out":3444,"would_cite":true,"duration_ms":40941,"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 closed-loop visuomotor policy trained entirely in simulation, using only optical flow from a wrist camera, guides a robot cutter to pruning points on real apple trees and orients the cutter perpendicular to the branch.","keywords":["robotic pruning","optical flow","sim-to-real transfer","visuomotor policy","reinforcement learning","synthetic tree generation","orchard robotics","L-systems"],"falsifier":"Count branch orientations, curvatures, and clutter densities in a set of real dormant V-Trellis Envy trees and compare them with the generator's output; if the real distribution puts many branches in the azimuth/elevation cells where the RL policy's success is near zero (around ±145 degrees azimuth at 0 degrees elevation) or if real trees have substantially more dense tertiary branching, the reported 30-35% success rates should be expected to drop in the field.","tokens_in":9529,"feed_emoji":"🍎","tokens_out":10376,"duration_ms":111553,"temperature":0.7,"pith_summary":"The paper tries to establish that the reaching and orientation part of dormant tree pruning can be done by a learned visuomotor policy that never sees a 3D reconstruction of the tree. The policy is trained in a purpose-built simulator that procedurally generates V-Trellis Envy apple trees from an L-system model with pruning and tie-down cycles, and it observes the world only as optical flow images from a wrist-mounted camera, plus the target cut point and robot state. In simulation the policy reaches correctly oriented cuts in about 30% of 3,000 trials, about half the 60% of an oracle planner with perfect geometry. On a physical proxy tree it achieves 35% success with zero-shot transfer, while a point-cloud-based RRT-Connect planner mostly fails because thin tertiary branches are missing from the reconstruction. The authors argue this makes reconstruction-free robotic pruning feasible and that optical flow is the key representation for seeing thin branches outdoors.","feed_headline":"Sim-trained robot policy reaches pruning cut points on real trees","feed_subtitle":"Uses wrist-camera optical flow, no 3D scan; 35% success on a real branch vs. planners missing thin branches.","key_machinery":"Two mechanisms carry the argument. The first is a procedural tree generator built on an L-system plant model extended with cyclic pruning and branch tie-down operations; branches are modeled as cantilever beams deflected to each tie point, which reproduces the curvature and clutter of V-Trellis Envy trees closely enough that a policy trained on them transfers to a physical tree. The second is the optical flow observation space: instead of RGB or depth, the policy receives a two-channel optical flow image computed between consecutive camera frames, which in field tests captures the geometry of thin tertiary branches better than a depth camera. Around these, a PPO-based LSTM policy with a privileged critic (given true branch orientation during training) and an autoencoding visual encoder learns the reaching, pointing, and perpendicularity behavior from shaped rewards.","core_discovery":"The authors claim that a closed-loop controller for the pruning reach task can be learned entirely in simulation and transferred zero-shot to a real robot by using optical flow images instead of RGB or depth. The controller takes a cut point and a stream of optical flow frames from an eye-in-hand camera, outputs 6-DOF end-effector velocities, and is rewarded for reducing distance to the cut point, pointing the cutter at the branch, and aligning the cutter jaws perpendicular to it. With a proximal-policy-optimization actor-critic trained on procedurally generated trees, the policy reaches the success criteria (jaws within 5 cm of the cut point, orientation within 30 degrees) in 30% of simulation trials and 35% of real-world trials, about half the simulation performance of an oracle RRT-Connect planner with perfect tree geometry and better than point-cloud-based planning on a real tree where thin branches are poorly reconstructed.","pith_inferences":["If the simulator's geometry is the bottleneck, a cheap validation step would be to measure branch-angle and curvature histograms from a few real trees and reject synthetic trees outside that distribution; this would make the zero-shot claim testable before field deployment.","The failure cells (branches pointing directly at or away from the robot) likely reflect the UR5's kinematic limits and the pruner's asymmetric mounting, so a different arm or a repositioning of the robot base could close part of the gap to the oracle.","Optical flow as an observation space may generalize to other manipulation tasks on thin, cluttered structures outside orchards, such as wiring, harvesting, or canopy inspection, wherever depth cameras are unreliable.","A direct extension would be to replace the point-cost specification with a learned detector that proposes cut points from the same optical flow stream, turning the reach policy into an end-to-end pruner."],"forward_implications":["Robotic pruning can proceed without a complete 3D model of the tree, sidestepping depth-sensor failures in bright sunlight and the cost of dense reconstruction.","Because optical flow preserves thin branches that point-cloud pipelines miss, the learned policy can reach cut points that a point-cloud-based planner cannot even see.","The policy's 30-35% success sits at roughly half of an oracle planner, so a practical system would still need a fallback planner or policy improvements, but it establishes a viable reconstruction-free baseline.","The success criteria are compatible with a contact-based admittance controller, so the learned reach-and-orient behavior can feed into an existing system that makes the actual cut once the branch is in the jaws.","The modified tree generator is not specific to V-Trellis; it can produce other planar orchard architectures (upright fruiting offshoot, tall spindle) with the same pruning-and-tie-down process."],"supporting_citations":[{"why":"Supplies the L-system plant model base that the authors extend with cyclic pruning and tie-down to generate synthetic orchard trees.","marker":"[12]"},{"why":"Establishes optical flow as a sim-to-real friendly representation that captures orchard branch geometry and motivates this paper's observation space.","marker":"[13]"},{"why":"Provides the off-the-shelf optical flow estimator used to compute the policy's two-channel flow images.","marker":"[21]"},{"why":"Represents the point-cloud-based vine pruning pipeline that motivates the classical planning baseline and its collision-relaxation trick.","marker":"[9]"},{"why":"Shows reinforcement learning can replace sampling-based planning for reaching vine pruning locations, the approach this paper extends to 6-DOF tree pruning.","marker":"[14]"},{"why":"Defines the contact-based admittance controller whose success criteria set the goal region for the learned reach policy.","marker":"[18]"},{"why":"Provides the model-free reinforcement learning algorithm used to train the policy.","marker":"[19]"},{"why":"Supplies the privileged-critic design that gives the critic true branch orientation during training only.","marker":"[25]"},{"why":"Provides the point-cloud registration library used to build the real-world tree reconstruction for the RRT-Connect comparison.","marker":"[28]"}],"fun_headline_variants":["Sim-trained pruning policy transfers zero-shot to real trees","Optical flow enables pruning robot without 3D scans","Sim-to-real pruning policy hits 35% on real trees","Wrist-camera optical flow guides pruning robot","Learned pruning controller reaches cut points on real trees"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the procedurally generated trees are geometrically close enough to real V-Trellis Envy trees that a policy trained on them will behave the same on a real tree; the paper validates this on a single proxy tree but has not formally checked the synthetic branch statistics against commercial orchards.","fun_headline_variants_meta":{"raw":{"variants":["Sim-trained pruning policy transfers zero-shot to real trees","Optical flow enables pruning robot without 3D scans","Sim-to-real pruning policy hits 35% on real trees","Wrist-camera optical flow guides pruning robot","Learned pruning controller reaches cut points on real trees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3363,"prompt_tokens":857,"completion_tokens":2506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2428}},"tokens_in":473,"tokens_out":2506,"duration_ms":22610,"temperature":1.0,"reasoning_tokens":2428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:07:42.761854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count branch orientations, curvatures, and clutter densities in a set of real dormant V-Trellis Envy trees and compare them with the generator's output; if the real distribution puts many branches in the azimuth/elevation cells where the RL policy's success is near zero (around ±145 degrees azimuth at 0 degrees elevation) or if real trees have substantially more dense tertiary branching, the reported 30-35% success rates should be expected to drop in the field.","supporting_citations":[{"cited_title":"L-systems: from the theory to visual models of plants,","cited_arxiv_id":null,"evidence_quote":"Supplies the L-system plant model base that the authors extend with cyclic pruning and tie-down to generate synthetic orchard trees."},{"cited_title":"Optical flow-based branch segmentation for complex orchard environments,","cited_arxiv_id":null,"evidence_quote":"Establishes optical flow as a sim-to-real friendly representation that captures orchard branch geometry and motivates this paper's observation space."},{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow,","cited_arxiv_id":null,"evidence_quote":"Provides the off-the-shelf optical flow estimator used to compute the policy's two-channel flow images."},{"cited_title":"Bumblebee: A Path Towards Fully Autonomous Robotic Vine Pruning","cited_arxiv_id":"2112.00291","evidence_quote":"Represents the point-cloud-based vine pruning pipeline that motivates the classical planning baseline and its collision-relaxation trick."},{"cited_title":"Reaching pruning locations in a vine using a deep reinforcement learning policy,","cited_arxiv_id":null,"evidence_quote":"Shows reinforcement learning can replace sampling-based planning for reaching vine pruning locations, the approach this paper extends to 6-DOF tree pruning."},{"cited_title":"Precision fruit tree pruning using a learned hybrid vision/interaction controller,","cited_arxiv_id":null,"evidence_quote":"Defines the contact-based admittance controller whose success criteria set the goal region for the learned reach policy."}],"review_version":1}