{"id":"aa095967-d680-460f-a22d-c706d1950b52","arxiv_id":"2509.18608","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Raw 3D LiDAR, compressed into flattened voxel maps, trains a reinforcement learning policy that reliably follows straight crop rows in simulation and degrades on curvier rows.","lead":"This paper trains a simulated agricultural robot to follow crop rows by feeding compressed 3D LiDAR scans into a reinforcement learning policy that outputs steering. It is a simulation-only proof of concept, with no tests on a real robot yet.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim-only results and the unvalidated fidelity of IsaacSim leave the central real-world row-following claim unsupported; Table III is environment-internal evidence.","rationale":"In good faith, the paper makes a modest, internally consistent simulation claim: a PPO policy over a 30x30 occupancy history can follow straight and mildly curved simulated rows. I checked the reported numbers and found no internal arithmetic error; the 95.83% compression figure is correct when the 7,200 points are counted as 21,600 coordinate values versus 900 row-map cells, though the wording is ambiguous. The main unresolved condition is the fidelity of IsaacSim to real under-canopy fields. The authors explicitly scope results to simulation (Section IV-B) and list real-world validation as future work (Section V), but Section IV-F's practicality statement and the introductory framing extend the claim beyond that scope. This is the same weakest assumption the reader identified, and the CONDITIONAL verdict is appropriate: accept only if the simulation is demonstrated to be faithful, or if the claim is explicitly restricted to simulation. I do not see a basis for REJECT, because the stated simulation results are internally coherent and the curved-row trajectory evidence shows the learned policy uses LiDAR to follow the row rather than driving open-loop.","tokens_in":8613,"tokens_out":9700,"duration_ms":78438,"concrete_test":"Re-run the Table III evaluation in IsaacSim with domain-randomized LiDAR noise/dropout, crop geometry (spacing, density, leaf area), and TerraSentia actuation parameters (e.g., ±30% variation), 15 trials per configuration. If the 100 m straight-row success rate drops below 100% or the 1.8 Hz/0.20 m mean distance drops below roughly 67 m (one standard deviation below the reported 83.89 m), the policy is brittle to plausible sim-to-real shifts and the real-world practicality claim is unsupported. Optionally, replay the policy on a real TerraSentia row-following dataset to test transfer directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an end-to-end LiDAR-based navigation policy for under-canopy agriculture, but all evidence comes from IsaacSim. Section IV-B asserts that the TerraSentia platform and plantation scenes were 'accurately modeled' and 'realistic,' yet no comparison to real LiDAR data, crop geometry, or platform dynamics is supplied. Section IV-F then extrapolates to real-world deployment ('real-world plantation layouts typically resemble the straight-row pattern'), while Section V admits real-world validation is future work. If simulator LiDAR returns, crop collision volumes/compliance, or TerraSentia dynamics diverge from real fields, the 100% straight-row success and the Table III distance numbers do not transfer. Because the abstract and introduction motivate the work by real-world GNSS and lighting failures, this unvalidated sim-fidelity premise is load-bearing, not merely a caveat. The internal simulation results are consistent and the curved-row trajectory in Fig. 5 shows perceptual control, so the concern is about external validity, not a numerical error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end crop-row navigation system for under-canopy agricultural robots. A PPO policy maps a compact LiDAR-derived representation directly to angular velocity commands (with constant forward speed) in IsaacSim. Raw Velodyne VLP-16 point clouds are voxelized and flattened across height channels into 30x30 row maps, and a three-step history of these maps is used as the policy input to address partial observability. The method is evaluated entirely in simulation: straight rows are completed with 100% success, curved rows trained at frequency 1.8 Hz and amplitude 0.20 m are followed for 83.89 +- 22.72 m on average, and performance degrades as frequency or amplitude moves away from the training distribution. Ablations show that removing the observation history or the downsampling pre-processing degrades learning.","tokens_in":8942,"tokens_out":8338,"duration_ms":71119,"significance":"If the central simulation claims hold, the paper is a useful demonstration that a compact voxel-flattened LiDAR representation, together with a small history buffer, is sufficient for a learned PPO policy to follow crop rows with mild curvature. The held-out frequency/amplitude tests are a genuine generalization check, and the 95.83% reduction in observation size is a practical contribution. The main weaknesses are external validity and reproducibility: the paper explicitly stops at simulation, yet contains a real-world deployment sentence, and a central reward parameter is unreported. The method is not compared to any existing row-following baseline. With corrections and tempered claims, the contribution would be a solid systems-oriented result.","major_comments":[{"comment":"The paper motivates the work by real-world GNSS and lighting failures, and Section IV-B asserts that the TerraSentia platform and plantation scenes were 'accurately modeled' and 'realistic.' No comparison to real LiDAR data, real crop geometry, or real platform dynamics is provided. Section IV-F then states that 'the practicality of the trained policy for real-world deployment' is highlighted by the straight-row result. Since all evidence is from IsaacSim and Section V lists real-world validation as future work, this deployment claim is unsupported. Please either remove/temper the deployment sentence or include a concrete sim-to-real check (e.g., a real-field point cloud or a qualitative platform trial).","section":"Section IV-B, IV-F"},{"comment":"The action-smoothing penalty is defined as clip(1 - ||omega_k - omega_{k-1}||_2 / sigma, [0,1]), but the scaling constant sigma is never reported. This is a free parameter of the reward that directly shapes the learned behaviors; without its value the training setup is not fully reproducible, and the sensitivity of the results to sigma is unknown. Please report sigma and, ideally, a small sensitivity study.","section":"Section III-C, Table I"},{"comment":"It is unclear whether the straight-row configuration is part of the training distribution or a held-out geometry. Section IV-D says policies were trained 'in a sinusoidal plantation environment,' yet Section IV-F calls the straight plantation a 'baseline configuration' and Figure 5 says both scenarios 'correspond to the training environment settings.' If straight rows were not seen in training, the 100% straight-row success is a generalization result and should be explicitly described as such; if they were seen, the claim should be qualified accordingly. This affects the interpretation of the abstract's headline result.","section":"Section IV-D, IV-F"},{"comment":"Eq. (5) is inconsistent with its verbal description. The text says the vertical axis is flattened by 'averaging the occupancy across all H height levels,' but the formula is o_{x,y} = (1/H) sum_{v_i} z_{v_i}, where z_{v_i} is the vertical voxel coordinate of a voxelized point. This computes an average z-coordinate (a height/depth map), not an average occupancy, and it double-counts voxels that contain multiple points. Please correct either the formula or the description; this is a central component of the observation pipeline and must be unambiguous.","section":"Section III-D, Eq. (5)"}],"minor_comments":[{"comment":"The paper repeatedly says the policy maps 'raw 3D LiDAR data' to control commands, but the input goes through a hand-specified voxelization, channel selection, and flattening step before the policy. Consider rephrasing to 'directly from 3D LiDAR with a fixed pre-processing transform' to avoid overstating the end-to-end claim.","section":"Sections I, III-D"},{"comment":"The training-return curves show single trajectories with no seeds or confidence intervals. The conclusion that 'without downsampling ... leads to failed learning' rests on one run. Please report multiple seeds or error bars, or state the number of seeds.","section":"Figure 6"},{"comment":"No comparison to existing row-following methods is provided. A geometric row detector (e.g., [8]) or a self-supervised baseline (e.g., [16]) would help the reader judge whether the learned end-to-end policy offers an actual advantage over the current state of the art, rather than just feasibility.","section":"Table III"},{"comment":"The 95.83% reduction figure is correct only if one counts each 3D point as three scalar values (7,200*3 = 21,600 values vs. 900 cells). If the intent is to compare point count to cell count, the reduction is 87.5%. Please state the basis for the claim.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The paper falls within scope for a robotics venue, but it reads as a work in progress. The simulation results are internally consistent, but the missing sigma, the Eq. (5) ambiguity, and the unclear training/evaluation split need to be fixed before publication. I would also encourage the editor to require at least one comparative baseline, and to ask the authors to either provide a minimal sim-to-real check or remove the real-world deployment sentence. The contribution is not yet at journal acceptance level."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a decent, internally consistent simulation study of an end-to-end PPO policy that maps raw 3D LiDAR to steering for row following, using a 30x30 flattened voxel map with a three-step history. The headline result—100% completion on straight 100m rows in IsaacSim, with graceful degradation as curvature increases—is believable on the numbers reported. What's new is the specific system integration: raw LiDAR through voxel downsampling into a compact BEV-like map as policy input, trained end-to-end in RL, with ablations showing both the downsampling and the history are necessary. That is application-level novelty, not a new algorithm, but it is a useful data point.\n\nThe best part is the ablation comparison. Without downsampling, learning fails; without history, the policy drifts and fails on curves. That supports the POMDP framing and makes the contribution concrete. The generalization sweep in Table III and Figures 7-8 is also a nice way to characterize robustness, and the declining performance is presented honestly.\n\nSoft spots, in order of importance. First, the central claim is only demonstrated in simulation, and the paper overreaches in Section IV-F with \"highlighting the practicality of the trained policy for real-world deployment.\" IsaacSim fidelity for crop geometry, LiDAR returns, and TerraSentia dynamics is asserted, not validated against real data. That's a load-bearing issue for the real-world motivation, though not for the internal sim claim. Second, there are no comparisons to existing row-following baselines, even simple ones like the heuristic LiDAR row detector from Higuti et al.—including one would calibrate how much the end-to-end approach actually buys. Third, a few reproducibility gaps: no seeds or variance for training curves in Figure 6, and the action-smoothing scale σ in Table I is never given. These are easy to fix. No code or assets are released either, which matters for RL papers.\n\nThe citation pattern is fine; the self-citations are related work, not load-bearing. I didn't find a numerical inconsistency in the main results. The paper deserves a serious referee rather than a desk reject—the internal sim study is sound as far as it goes—but the revision needs baselines, reproducibility details, and a much more careful statement of what is and isn't validated.\n\nMy recommendation: send it to peer review, expect heavy revision. It is useful for researchers working on RL for under-canopy navigation, but I would not cite it as evidence for real-world deployment.","headline":"Solid sim-only proof-of-concept for end-to-end LiDAR RL row following; the sim-to-real jump is not supported.","tokens_in":9331,"tokens_out":2221,"would_cite":true,"duration_ms":17135,"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":"An end-to-end reinforcement learning policy can steer an agricultural robot down crop rows using only raw 3D LiDAR, trained entirely in simulation, with 100% success on straight rows and gradual decline on curved rows.","keywords":["crop row navigation","under-canopy agriculture","deep reinforcement learning","LiDAR perception","voxel downsampling","row following","sim-to-real transfer","autonomous agricultural robot"],"falsifier":"Run the trained policy on the same physical robot in a real under-canopy field with straight and mildly curved rows. If it collides or veers out of the row within the first 10 meters while simulation rollouts succeed, the central claim fails as stated; a quantitative check would compare real distance-before-collision against the simulated 100 m straight-row and 83.89 m curved-row baselines.","tokens_in":8552,"feed_emoji":"🌱","tokens_out":7211,"duration_ms":51861,"temperature":0.7,"pith_summary":"This paper seeks to establish that an under-canopy agricultural robot can learn to follow crop rows end-to-end: raw 3D LiDAR point clouds go directly into a reinforcement learning policy that outputs steering commands, with no labeled datasets, no hand-crafted row-detection heuristics, and no separate perception-to-control interface. The enabler is a voxel-based downsampling step that turns a point cloud into a compact 2D row map, cutting the input by 95.83%, and stacking three consecutive maps so the policy has temporal context. In simulation, the policy completes 100-meter straight rows in every trial and traverses curved rows with gradually decreasing distance as sinusoidal frequency or amplitude increases. The authors take this as evidence that row-following can be achieved without GPS, without cameras, and without manual tuning, pending real-world validation.","feed_headline":"100% success on straight rows for sim-trained LiDAR policy","feed_subtitle":"A 30x30 voxel map cuts LiDAR input 95.83%; a three-frame history lets the policy follow rows without labels or GPS.","key_machinery":"The central object is a compact row map: raw LiDAR points are voxelized at 0.1 m resolution and vertically flattened by averaging occupancy across the middle four LiDAR channels, yielding a 30x30 grid of 900 cells. A history of three consecutive row maps is the policy input, turning the partially observable problem into a fully observable approximation. The map is what makes learning feasible: the raw 7,200-point cloud overwhelms the network, while the compact map with temporal context trains successfully.","core_discovery":"The central claim is that a policy trained with proximal policy optimization, fed a history of three 30x30 voxel-flattened LiDAR row maps and rewarded for forward progress while penalized for collisions and abrupt steering, learns a row-following controller that generalizes within the training distribution. The evidence is entirely simulated: 100% success over 100 m on straight rows; 83.89 m average on the curved training configuration; and monotonic performance decline to 9.46 m at the highest tested frequency and 19.57 m at the highest tested amplitude. The authors argue this supports treating the partially observable problem as a fully observable Markov decision process with an observatio","pith_inferences":["Because the policy reads only local row geometry, the same voxel-map architecture could extend to other structured vegetated corridors such as orchards and vineyards, though the paper does not test this.","The three-frame history appears to carry useful velocity information, since the single-frame policy fails on curves; a controlled study comparing history lengths would isolate how much temporal context contributes.","A direct sim-to-real deployment with the same voxel map and no retraining would separate simulation-fidelity issues from policy limitations; the paper explicitly leaves real-world validation to future work.","The monotone performance decline across frequency and amplitude could serve as a calibration curve: matching that decline on real fields would validate the simulator, and divergence would localize the sim-to-real gap."],"forward_implications":["A LiDAR-only row-following controller can be trained entirely in simulation, eliminating labeled field data and manually designed perception-to-control interfaces.","The 95.83% input reduction makes a 7,200-point cloud a 900-cell map that a deep reinforcement learning policy can master within the 2 million environment steps shown in training.","Straight-row plantations, the most common real-world layout, are navigable with 100% success over 100 m in simulation.","Performance degrades gradually rather than abruptly as row curvature increases, so the policy retains partial utility under moderate unseen curvature.","The reward design—forward progress, collision penalty, and a clipped action-smoothness penalty—is sufficient to keep the robot near the row centerline without an explicit lateral-error term."],"fun_headline_variants":["Sim-only RL policy hits 100% success on straight rows","Sim-trained LiDAR policy navigates rows without GPS or labels","Voxel maps cut LiDAR data 95.83% for RL row following","Deep RL row navigation: straight rows perfect, curves degrade"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the simulated crops, LiDAR returns, and robot dynamics faithfully reproduce real under-canopy fields; all results are measured inside that simulator, and real-world validation is left to future work.","fun_headline_variants_meta":{"raw":{"variants":["Sim-only RL policy hits 100% success on straight rows","Sim-trained LiDAR policy navigates rows without GPS or labels","Voxel maps cut LiDAR data 95.83% for RL row following","Deep RL row navigation: straight rows perfect, curves degrade"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1063,"prompt_tokens":649,"completion_tokens":414,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":393,"completion_tokens_details":{"reasoning_tokens":339}},"tokens_in":393,"tokens_out":414,"duration_ms":3784,"temperature":1.0,"reasoning_tokens":339,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:41:33.298544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy on the same physical robot in a real under-canopy field with straight and mildly curved rows. If it collides or veers out of the row within the first 10 meters while simulation rollouts succeed, the central claim fails as stated; a quantitative check would compare real distance-before-collision against the simulated 100 m straight-row and 83.89 m curved-row baselines.","supporting_citations":[],"review_version":1}