Pith. sign in

REVIEW 4 major objections 4 minor 20 references

A drone that learns to efficiently find non-uniformly distributed objects in agricultural fields: from simulation to the real world

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A drone path planner trained entirely in an abstract grid simulator transfers to real agricultural fields, finding clustered objects with about 72% shorter flight paths than a full coverage planner while giving up about 25% recall.

desk verdict A useful sim-to-real case study with an honest limitation, but the headline numbers need reconciling and the evidence base is a single field. read the letter →

arxiv 2505.09278 v2 pith:WCJEGQMU submitted 2025-05-14 cs.RO

Rick van Essen , Eldert van Henten , Gert Kootstra This is my paper · ORCID
classification cs.RO
keywords reinforcementlearningUAVpathplanningprecisionagricultureweeddetectionsim-to-realtransferadaptivepriorknowledgecoverage
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a drone can learn, entirely in an abstract grid simulation, to search an agricultural field for non-uniformly distributed objects such as weed patches, and then fly that policy on a real drone. The learned agent combines low-quality prior knowledge from a quick high-altitude pass with live detections to decide where to fly and when to land. Measured across four levels of realism, the policy shortens flight paths by roughly three-quarters in simulation and by 72% on real-world orthomosaics, at the cost of 14% and 25% lower recall than a full coverage planner. The point of the exercise is that many monitoring tasks tolerate missed objects, so a large flight-time saving can be worth a moderate drop in recall; the paper quantifies where the simulation-to-reality gap comes from.

What carries the argument

The load-bearing mechanism is the Markov Decision Process and its drone-centered state representation. The agent sees a global map (field borders, already-found objects, coverage, and the uncertain prior-knowledge map) and a local field-of-view map of current detections, plus the battery level; from this a Deep Q-Network outputs one of five actions (fly north/south/east/west or land). The reward design—positive reward per discovered object, penalties per step, for revisiting seen cells, and for landing too late—is what makes the policy prefer cluster-seeking shortcuts over exhaustive coverage. The four-level evaluation protocol (simulation, real object positions, real prior knowledge, full real-world imagery) is the instrument that isolates how much each sim-to-real mismatch costs.

What would settle it

Take the trained agent to a field where objects are spread uniformly rather than clustered, with the same prior-knowledge generation procedure, and measure recall at 400 flight steps against the full coverage planner; the paper's mechanism predicts the RL agent still finds far more objects early. If its early recall is not above the coverage planner's, the claim that the policy exploits clumped distributions fails.

Watch

Extended reading notes

Core claim

The central claim is that a Deep Q-Network trained on an abstract MDP—a square grid with Gaussian object clusters, hand-set detection and prior-knowledge error probabilities, and rewards for detections versus flight steps—produces a flight policy that transfers to real agricultural search. On four real-world orthomosaic datasets with clustered artificial plants, the agent flew about 269 steps versus 956 for a full coverage planner (a 72% shorter path) and reached 0.70 recall versus 0.95 for full coverage; in full simulation it reached 0.86 recall in about 242 steps. Decomposing the gap by replacing simulator components one at a time, the paper attributes about one-third of the real-world recall loss to the different object distribution and about two-thirds to the lower quality of real prior knowledge, with the real detector being less error-prone than the simulated error model assumed. The authors conclude that for applications where finding all objects is not essential, such as weed detection, the learned planner is suitable and efficient.

Load-bearing premise

The whole transfer rests on the training simulator's model of the real world—especially that object clusters are spread across the whole field and that prior knowledge is only moderately unreliable; when those assumptions miss, real-world recall drops by more than the simulation predicted.

Editorial extensions

If this is right

  • In monitoring tasks where missing some objects is acceptable, the learned planner can cover a field in roughly a quarter of the flight time, extending the area a battery-limited drone can inspect.
  • Improving prior knowledge—better alignment between grid cells and real positions, or training with noisier priors—should recover a large share of the real-world recall gap, since two-thirds of it is attributed to prior quality rather than to the detector.
  • The agent's early-flight advantage is pronounced: at 200 flight steps on real-world data it already finds around 64% of objects versus 23% for the coverage planner, so time-critical scouting can get most information very early.
  • For applications that require finding every object, full coverage or the rule-based planner remains more appropriate; the learned planner is positioned on the speed-versus-recall trade-off.
  • Repeated monitoring flights can compensate for missed weeds: objects missed in one flight may be found in a later one, making the recall loss less consequential over a season.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors' own observation that outer objects in each cluster are missed suggests a simple extension they do not test: adding a cluster-completion reward or a penalty for leaving a partially explored cluster would likely recover recall at modest flight-time cost.
  • An implicit generalization is that the same simulator-to-real recipe—abstract grid MDP, uncertain prior map, and a light detector—should transfer to other patchy agricultural phenomena, such as disease foci or nutrient-deficiency zones, wherever the cost of a false negative is low.
  • Because the largest identified loss comes from prior-knowledge quality, a testable design choice is to train with prior maps whose false-negative probability is much higher than the current 0.25 and whose spatial alignment errors are larger; the paper's own level-3 results predict this would harden the policy against real priors.
  • The paper evaluates on a rectangular field mapped into a square simulator; training on variable aspect ratios and cluster placements that do not fill the whole area could close roughly the one-third of the gap attributed to object distribution.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents a reinforcement-learning-based adaptive path planner for drone search in agricultural fields, trained in an abstract grid simulator that models object distributions, detection errors, and uncertain prior knowledge. The agent is evaluated at four levels of realism: pure simulation, simulation with real object positions, simulation with real prior knowledge, and offline real-world orthomosaic data. The authors report large flight-path reductions relative to a full coverage planner (e.g., 78% shorter in simulation and 72% shorter on real-world data) at the cost of lower recall, and conclude that the approach is suitable for applications such as weed detection that tolerate false negatives. The central claim is that a purely simulation-trained policy transfers to real-world agricultural search with a substantial flight-time saving at a moderate recall cost.

Significance. If the results hold, the paper provides a rare end-to-end demonstration of sim-to-real transfer for an RL-based adaptive search planner using uncertain prior knowledge, and the four-level evaluation design is a thoughtful way to decompose the simulation-to-reality gap. The comparison against a standard coverage baseline and the attribution of the recall gap to prior-knowledge quality and object distribution are useful contributions. However, the evidence for generalization is narrow, the headline numbers are internally inconsistent, and the simulation parameters do not match the measured real-world perception noise, so the central quantitative claim is not yet established at archival standard.

major comments (4)
  1. [Abstract vs. Section III/Table IV and Section V] The headline quantitative results are inconsistent across the manuscript. The abstract reports a 57% shorter flight path in simulation with 13% lower recall and a 38% shorter flight path on real-world orthomosaic data with 21% lower recall, while Section III and Table IV report 78% shorter with 14% lower recall and 72% shorter with 25% lower recall, and the Conclusion states 75% shorter in simulation with 14% lower recall and 72% shorter on real-world data with 25% lower recall. The abstract also reports real-world drone flight trials finding 73% and 23% of objects, but the body contains no such trials. These discrepancies directly affect the paper's headline contribution and must be reconciled with the underlying data.
  2. [Section II-E and II-F] The sim-to-real transfer claim rests on a single field with artificial plants: the four real-world datasets in Section II-E are orthomosaics of one grass field, and the YOLOv8-nano detector in Section II-F was trained on images taken from that same field. Because the detector feeds both the observation map and the prior knowledge map (Section II-D.2-II-D.4), the Level 4 results do not evaluate the RL policy independently of this particular perception pipeline. A stronger test, such as a second field with different geometry or plant type and a detector not trained on the evaluation field, is needed before claiming general transfer to agricultural fields.
  3. [Table I and Section III] The simulation error parameters are inconsistent with the measured real-world perception quality, which weakens the decomposition of the sim-to-real gap. Table I sets the detection false-negative probability p_dt_fn to 0.0001, yet the detection network in Section II-F has a validation recall of 0.95, implying a false-negative rate of about 0.05. Similarly, the simulated prior knowledge has recall 0.42±0.06 (Section III) versus 0.16±0.04 for real prior knowledge. The one-third/two-thirds attribution in Section IV is therefore computed with a simulator that substantially underestimates detection noise and overestimates prior-knowledge quality, so the conclusions about which factors matter for transfer are not yet robust.
  4. [Section IV, flight path length accounting] The reported flight-path savings exclude the high-altitude full-coverage flight needed to generate the prior knowledge. Section IV states that including this flight would add about 250 flight steps, and although the authors argue the RL agent would still beat the coverage planner, the headline 72% saving (Table IV, Level 4) does not include this required overhead. The paper should report both with and without the prior-knowledge flight so readers can compare the true end-to-end flight cost.
minor comments (4)
  1. [Table I] The second covariance matrix row is labeled 'Σ1' but should be 'Σ2'; this is likely a typographical error that should be fixed.
  2. [Section II-F] The subsection heading 'F . Training of the detection network' contains a stray space between 'F' and the period; please correct the formatting.
  3. [Abstract] The abstract's 57%/38% savings and 13%/21% recall numbers should be harmonized with the body, since the mismatch can mislead readers who only consult the abstract.
  4. [Section II-D.4] The confidence threshold for prior knowledge detections is set to 0.05, while the detection network threshold is 0.5; the rationale for this asymmetry and its effect on prior-knowledge precision/recall should be stated explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the RL policy is trained on randomized simulations and evaluated on separate real-world orthomosaics; the reported trade-off is measured, not fitted.

full rationale

The paper's derivation chain is not circular. The DQN policy is trained in an abstract simulator with randomly generated fields (Sec. II-A, Table I) and is then evaluated on four real-world orthomosaics (Level 4, Sec. II-H). No parameter of the policy is fitted to the Level 4 data, and the reported 72% shorter flight path at 25% lower recall is an empirical measurement against an independent full-coverage baseline (Fields2Cover), not a quantity reconstructed from a fitted input. The simulation parameters are hand-chosen modeling choices, and the paper explicitly attributes part of the real-world performance gap to mismatches in object distribution and prior-knowledge quality (Sec. IV), which is the opposite of disguising inputs as outputs. Self-citations to the authors' prior work [12] and [18] supply the simulator, training details, the dataset, and a rule-based comparison, but these are incremental-reuse citations rather than load-bearing justifications of the transfer claim; the core evaluation rests on external orthomosaic data and an external coverage planner. One legitimate weakness, the detection network being trained on images from the same grass field used in Level 4, affects generalization but does not make the planner comparison circular, since the planner itself never sees real labels during training. Overall, the central claim has independent empirical content, with only minor non-load-bearing self-citation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a simulation whose parameters (object counts, cluster shapes, error rates, reward weights) are chosen by hand and, as the paper's own discussion shows, are more optimistic than real-world conditions. No new physical entities are introduced.

free parameters (6)
  • Object count distribution N_obj(mu=100, sigma=30) = 100 +/- 30
    Sets the number of target objects in the field; directly influences path-length and recall trade-offs. Chosen by hand for the simulation, not measured from the real datasets.
  • Cluster count distribution N_k(mu=5, sigma=2) = 5 +/- 2
    Determines how clustered the objects are; the real field had a different, less field-filling distribution, and the paper attributes one-third of the sim-to-real recall drop to this.
  • Covariance matrices Sigma_1 and Sigma_2 = [5 8;8 15] and [15 0;0 5]
    Control cluster shape and orientation; chosen by hand; no evidence they match real plant clusters.
  • Prior-knowledge miss rate p_pk_fn = 0.25
    Governs how incomplete the prior knowledge is. The real-world prior knowledge achieved recall 0.16, far below the simulated prior recall 0.42, so the training scenario is more optimistic than reality.
  • Detection miss rate p_dt_fn = 0.0001
    Almost no false negatives in simulated detections, while the real YOLOv8 detector has recall 0.95 overall; the simulator's detection model is far more optimistic than the real detector.
  • Reward terms (r_dt, r_step, r_nocov, r_nfz, r_crash) = 1.0, -0.125, -0.5, -1.0, -150.0
    Reward shaping constants that define the trade-off between finding objects and path length; the paper adjusted b_init and r_step relative to [12] for the larger field.
assumptions (4)
  • domain assumption The drone-centered global and local binary maps, plus battery level, form a Markovian state sufficient for optimal search decisions.
    This is the MDP state representation (Sec. II-B1). If the low-resolution global map or the coverage layer discards task-relevant structure, the learned policy is suboptimal.
  • domain assumption Object positions and prior-knowledge errors are independent and can be modeled as Gaussian clusters and Bernoulli pixel flips.
    Used throughout the training simulator (Sec. II-A). The real-world data violate this: field geometry is rectangular, clusters do not fill the square field, and prior-knowledge errors are spatially correlated through misregistration at cluster boundaries.
  • ad hoc to paper The reward function correctly encodes the mission objective (object count vs flight length vs crash avoidance).
    Rewards in Table I are chosen by hand. A different weighting, e.g., higher r_dt, would produce longer flights and higher recall, as the authors note in the discussion.
  • domain assumption The YOLOv8-nano detector's outputs (after threshold 0.5) are an unbiased observation model for the simulation's binary detection map.
    The real detector is trained on the same field and may have spatially varying performance; the simulation treats detections as independent Bernoulli events, which is unlikely to hold in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A drone that learns to efficiently find non-uniformly distributed objects in agricultural fields: from simulation to the real world." pith.science (2026). https://pith.science/paper/WCJEGQMU

@misc{pith2026250509278,
  author       = {Pith},
  title        = {Pith review of: A drone that learns to efficiently find non-uniformly distributed objects in agricultural fields: from simulation to the real world},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WCJEGQMU}},
  note         = {Machine review of arXiv:2505.09278}
}
read the original abstract

Drones are promising for data collection in precision agriculture but are limited by battery capacity. Drone paths are usually planned using full coverage planners, even though this is not always required. This paper presents a drone path planner trained with Reinforcement Learning (RL) to detect as many objects as possible with a minimal flight path length. The agent uses low-quality prior knowledge derived from a high-altitude full coverage flight as guidance. The agent was trained in simulation, modeling object distributions, drone movement, field geometry, detection errors, and uncertain prior knowledge. Combined with a flight controller and object-detection network, it controls flight direction, terminates flights, and can be deployed on a real drone. It was evaluated across six levels of realism, from pure simulation to real-world drone flights, to quantify the simulation-to-reality gap. The agent achieved a 57% shorter flight path than a full coverage planner in simulation (13% lower recall) and a 38% shorter flight path on real-world orthomosaic data (21% lower recall). In real-world drone flights, the agent found 73% and 23% of the objects in trials 1 and 2, respectively; the lower real-world performance was mainly attributed to prior knowledge quality. Although framed as a weed-detection task, the approach is expected to generalize to other agricultural tasks with non-uniformly distributed objects and tolerance to false negatives, though further research is needed before practical deployment.

Figures

Figures reproduced from arXiv: 2505.09278 by the authors.

Figure 1
Figure 1. Overview of the learning-based adaptive path planner, which learns a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DQN architecture, showing the pooling, convolutional, and fully [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Mean reward of the evaluation simulations during training of the DQN [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Example of the artificial plants in the grass field. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 5
Figure 5. Figure 5: Effect of flight path length on the recall for the RL agent (DQN) and [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Flight paths for the RL agent on the four real-world datasets. Blue, orange, and red markers indicate true positive, false positive, and false negative [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

  1. [12]

    UAV-based path planning for efficient localization of non-uniformly distributed weeds using prior knowledge: A reinforcement-learning approach

    R. van Essen, E. van Henten, and G. Kootstra, “Learning UA V-based path planning for efficient localization of objects using prior knowledge,” 2024, arXiv:2412.11717 [cs]

    work page Pith review arXiv 2024

  2. [1]

    Drones in Precision Agri- culture: A Comprehensive Review of Applications, Technologies, and Challenges,

    R. Guebsi, S. Mami, and K. Chokmani, “Drones in Precision Agri- culture: A Comprehensive Review of Applications, Technologies, and Challenges,”Drones, vol. 8, no. 11, p. 686, 2024

    work page 2024

  3. [2]

    A survey on deep learning-based identification of plant and crop diseases from UA V-based aerial images,

    A. Bouguettaya, H. Zarzour, A. Kechida, and A. M. Taberkit, “A survey on deep learning-based identification of plant and crop diseases from UA V-based aerial images,”Cluster Computing, vol. 26, no. 2, pp. 1297– 1317, 2023

    work page 2023

  4. [3]

    An automatic method for weed mapping in oat fields based on UA V imagery,

    M. Gašparovi ´c, M. Zrinjski, Ð. Barkovi´c, and D. Radoˇcaj, “An automatic method for weed mapping in oat fields based on UA V imagery,” Computers and Electronics in Agriculture, vol. 173, p. 105385, 2020

    work page 2020

  5. [4]

    Can Basic Soil Quality Indicators and Topography Explain the Spatial Variability in Agricultural Fields Observed from Drone Orthomosaics?

    R. Näsi, H. Mikkola, E. Honkavaara, N. Koivumäki, R. A. Oliveira, P. Peltonen-Sainio, N.-S. Keijälä, M. Änäkkälä, L. Arkkola, and L. Alakukku, “Can Basic Soil Quality Indicators and Topography Explain the Spatial Variability in Agricultural Fields Observed from Drone Orthomosaics?”Agronomy, vol. 13, no. 3, p. 669, 2023

    work page 2023

  6. [5]

    Drones in agri- culture: A review and bibliometric analysis,

    A. Rejeb, A. Abdollahi, K. Rejeb, and H. Treiblmaier, “Drones in agri- culture: A review and bibliometric analysis,”Computers and Electronics in Agriculture, vol. 198, p. 107017, 2022

    work page 2022

  7. [6]

    Extent and Implications of Weed Spatial Variability in Arable Crop Fields,

    S. Garibay, W. Richner, P. Stamp, T. Nakamoto, J. Yamagishi, C. Abi- vardi, and P. Edwards, “Extent and Implications of Weed Spatial Variability in Arable Crop Fields,”Plant Production Science, vol. 4, no. 4, pp. 259–269, 2001

    work page 2001

  8. [7]

    The Spatial Analysis of Soilborne Pathogens and Root Diseases,

    C. L. Campbell and J. P. Noe, “The Spatial Analysis of Soilborne Pathogens and Root Diseases,”Annual Review of Phytopathology, vol. 23, no. 1, pp. 129–148, 1985

    work page 1985

Show all 20 references
  1. [8]

    Learning- based methods for adaptive informative path planning,

    M. Popovi ´c, J. Ott, J. Rückin, and M. J. Kochenderfer, “Learning- based methods for adaptive informative path planning,”Robotics and Autonomous Systems, vol. 179, p. 104727, 2024

    2024

  2. [9]

    R. S. Sutton and A. G. Barto,Reinforcement learning: an introduction. Cambridge, Massachusetts, USA: MIT press, 2018

    2018

  3. [10]

    Drone Deep Reinforcement Learning: A Review,

    A. T. Azar, A. Koubaa, N. Ali Mohamed, H. A. Ibrahim, Z. F. Ibrahim, M. Kazim, A. Ammar, B. Benjdira, A. M. Khamis, I. A. Hameed, and G. Casalino, “Drone Deep Reinforcement Learning: A Review,” Electronics, vol. 10, no. 9, p. 999, 2021

    2021

  4. [11]

    UA V Path Planning and Obstacle Avoidance Based on Reinforcement Learning in 3D Environments,

    G. T. Tu and J. G. Juang, “UA V Path Planning and Obstacle Avoidance Based on Reinforcement Learning in 3D Environments,”Actuators 2023, V ol. 12, Page 57, vol. 12, no. 2, p. 57, 2023, publisher: Multidisciplinary Digital Publishing Institute

    2023

  5. [13]

    Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Survey,

    W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Survey,” in2020 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, Dec. 2020, pp. 737–744

    2020

  6. [14]

    UA V Path Planning using Global and Local Map Information with Deep Reinforcement Learning,

    M. Theile, H. Bayerlein, R. Nai, D. Gesbert, and M. Caccamo, “UA V Path Planning using Global and Local Map Information with Deep Reinforcement Learning,” in2021 20th International Conference on Advanced Robotics (ICAR). Ljubljana, Slovenia: IEEE, 2021, pp. 539– 546

    2021

  7. [15]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep...

    2015

  8. [16]

    Ultralytics YOLO,

    G. Jocher, J. Qiu, and A. Chaurasia, “Ultralytics YOLO,” 2023, version: 8.3.49. [Online]. Available: https://github.com/ultralytics/ultralytics

    2023

  9. [17]

    Metashape Professional,

    Agisoft, “Metashape Professional,” St. Petersburg, Russia, 2023, version: 2.0.3. [Online]. Available: https://www.agisoft.com

    2023

  10. [18]

    Adaptive path planning for efficient object search by UA Vs in agricultural fields,

    R. van Essen, E. van Henten, L. Kooistra, and G. Kootstra, “Adaptive path planning for efficient object search by UA Vs in agricultural fields,” 2025, arXiv:2504.02473 [cs]

    2025 arXiv

  11. [19]

    Fields2Cover: An open-source coverage path planning library for unmanned agricultural vehicles,

    G. Mier, J. Valente, and S. de Bruin, “Fields2Cover: An open-source coverage path planning library for unmanned agricultural vehicles,” IEEE Robotics and Automation Letters, 2023

    2023

  12. [20]

    A Performance Analysis of You Only Look Once Models for Deployment on Constrained Computational Edge Devices in Drone Applications,

    L. Rey, A. M. Bernardos, A. D. Dobrzycki, D. Carramiñana, L. Bergesio, J. A. Besada, and J. R. Casar, “A Performance Analysis of You Only Look Once Models for Deployment on Constrained Computational Edge Devices in Drone Applications,”Electronics, vol. 14, no. 3, p. 638, Feb. ...

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.