Pith. sign in

REVIEW 4 major objections 4 minor 47 references

A zero-shot vision-language model can run a crop-monitoring robot through short tasks as well as humans do, but its performance collapses on long multi-plant, multi-target runs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 09:55 UTC pith:OT2W6ZZH

load-bearing objection Useful new benchmark and an honest negative result for VLM long-horizon planning, but the near-human claim rests on oracle maps and human stopping, so the paper overstates deployment readiness. the 4 major comments →

arxiv 2601.11906 v2 pith:OT2W6ZZH submitted 2026-01-17 cs.RO

Visual-Language-Guided Task Planning for Horticultural Robots

Jose Cuaran , Kendall Koe , Aditya Potnis , Naveen Kumar Uppalapati , Girish Chowdhary This is my paper
classification cs.RO
keywords crop monitoringtask planningvision-language modelshorticultural roboticssemantic occupancy mapmobile manipulationzero-shotagricultural robotics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether a vision-language model can act as the high-level planner for a mobile manipulator doing crop monitoring in greenhouses. The authors build a framework in which the model interleaves queries to a semantic occupancy map and camera feeds with navigation and manipulation commands, and they benchmark it on short- and long-horizon tasks in simulated mono- and polyculture scenes. They report that on the simplest category—one plant, one target—the zero-shot model succeeds 86.7% of the time, close to the 92% human rate, while on multi-plant, multi-target tasks success falls to 9.38%. The paper also shows that under the system's own noisy semantic map, success drops to roughly 39–46%, exposing perception noise and long-context grounding as the main constraints.

Core claim

The central claim, stated in the authors' own terms, is that a vision-language model can plan and execute short-horizon crop monitoring tasks in a zero-shot fashion at a level approaching human operators, provided it is given ideal map information and human-supervised stopping. The evidence is the 86.7% versus 92% success on single-plant single-target tasks across three environments, with 100% in the most complex scene. The same system, however, drops to 42.11% success on multi-plant single-target tasks and 9.38% on multi-plant multi-target tasks, even though task completion rates stay above 76%. The authors attribute the collapse to weak spatial reasoning, context-window limits, and repetit

What carries the argument

The load-bearing mechanism is an iterative tool-calling agent: a vision-language model that, at each step, chooses among perception tools (a top-down semantic occupancy map, a robot-centric map, base camera with overlaid polar action commands, tip camera with a 2D grid) and action tools (global navigation, local moves, manipulator primitives, a centering tool). Instead of generating a full plan upfront, the model acts one step at a time and replans from feedback, which suits partially observable crops. The semantic map is built from open-vocabulary detections back-projected and filtered into a sparse object map fused with an occupancy grid; this map supplies the spatial context the model rea

Load-bearing premise

The near-human headline result assumes the robot is given ground-truth object locations and a human decides when each task is finished; when the robot's own noisy semantic map is used instead, success drops to about 39–46%.

What would settle it

Run the single-plant, single-target benchmark without ground-truth object locations and without human stopping, using only the robot's own noisy map and an automatic or learned success check. If success on that simplest category falls to the 40–50% range seen in the ablations, the claim that zero-shot VLM planning is near-human for simple crop monitoring would be falsified for practical deployment.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Short-horizon monitoring—go to one plant and image one target—can be delegated to a zero-shot VLM with near-human success under clean map conditions.
  • Long-horizon, multi-plant, multi-target tasks remain unsolved: success below 10%, and even partial completion requires many more corrective tool calls than human operation.
  • Perception quality is decisive: substituting the noisy semantic map for ground-truth locations cuts success by about 12 percentage points and completion by about 37 points, so improving object detection and map fusion directly improves planning.
  • Text-only few-shot demonstrations do not recover the lost performance; the agent using the front camera with polar-action overlays outperforms the agent using a robot-centric top-down map, suggesting VLMs reason better in image space.
  • VLM control is reliably less efficient than human teleoperation, measured both by tool-call counts and by path-efficiency relative to humans, which will matter for battery- and time-constrained field robots.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The sharp cliff between one-target and multi-target tasks suggests context length and memory, not object recognition, is the dominant bottleneck; a testable extension is to fix the number of plants while varying the number of targets and measuring success decay.
  • If VLMs reason better from egocentric camera overlays than from top-down maps, then a practical direction is a hybrid: use maps for global planning and camera-space prompts for local control, or fine-tune small VLMs on map-like inputs.
  • The reliance on human stopping in the headline numbers implies that an autonomous success detector could become a component worth benchmarking separately; replacing human supervision with a learned success detector would likely lower the headline numbers further.
  • The simulated scenes and real-time mapping pipeline suggest a direct field test in a greenhouse: running the same single-plant task with real noisy detections would tell whether the 87% result survives the sim-to-real gap.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a modular framework in which a vision-language model (GPT-4.1) controls a simulated mobile manipulator for horticultural crop-monitoring tasks. The agent iteratively selects perception and action primitives—querying semantic occupancy maps, robot-centric maps, and RGB camera images—to navigate to plants and capture close-up views of targets. The authors evaluate the system on short- and long-horizon tasks in three simulated greenhouse environments, comparing zero-shot VLM control against human teleoperators, and report ablations with noisy semantic maps and few-shot prompting. Their headline claim is that zero-shot VLM control approaches human-level success (86.7% vs. 92%) on single-plant/single-target tasks, while performance collapses to 9.38% on multi-plant/multi-target tasks. They identify map noise, long-horizon context management, and spatial reasoning as key limitations.

Significance. If interpreted as an upper-bound study of VLM planning under idealized perception and human oversight, this is a useful contribution: it provides a new simulated benchmark, a modular tool-calling design, a detailed failure-mode taxonomy, and a clean demonstration that zero-shot VLMs can handle short-horizon agricultural tasks but not long-horizon ones. The benchmark and ablation methodology are potentially reusable by the community. However, the headline near-human claim conflates the VLM planner with an oracle map and human success detection; the paper's own ablations show large performance drops (to 39–46% average success) under noisy maps. The central claim therefore needs re-framing rather than rejection, and the paper would be strengthened by explicitly separating 'VLM planning with ideal perception and human supervision' from 'autonomous VLM-controlled system.'

major comments (4)
  1. [Section IV.C and Result 2 (Table I)] The 86.7% success rate for single-plant/single-target tasks is not attributable to the VLM alone: the setup provides ground-truth object locations ('we provide the ground-truth object locations to analyze the VLM's decision-making performance under ideal conditions') and human experts manually stop execution when the task is complete. Since Failure Mode 3 documents that the VLM frequently mis-evaluates its own success, human stopping masks these errors and inflates success. The noisy-map ablations (Tables V–VI) still use human stopping, so they do not isolate autonomous success detection either. To support the 'approaching human performance' claim, the authors should either (a) re-run without oracle maps and with automated success detection, or (b) explicitly re-label the result as 'VLM planning under oracle perception and human oversight.' The abstract and conclusion currently overstate
  2. [Section V.A, task definition] The success criterion treats a missing target attribute (e.g., a green or yellow tomato absent on the selected plant) as a failure, and the authors acknowledge this is an artifact of the task definition. This artifact differentially affects environments and operators: for example, human single-plant/single-target success is 100% in the complex polyculture but only 85.71% in the simple polyculture. Because the reported success rates therefore reflect environmental/task-design factors as well as planning ability, the human–VLM comparison is not a clean capability comparison. The artifact should be removed (e.g., by sampling only plants known to contain the target) or its incidence should be quantified and reported separately for each condition.
  3. [Section IV.B, SPL definition] The metric is named 'success rate weighted by shortest path length' and cites [45], but l_i is defined as the distance traveled by the human expert, not the shortest path length. This makes the reported SPL a measure of efficiency relative to human routes, not relative to optimal paths. Human routes may themselves be suboptimal, so the SPL values in Tables I–VI are not directly comparable to standard SPL results in the navigation literature. Please rename the metric (e.g., 'success weighted by human-path ratio'), clarify the difference from [45], and discuss how this choice affects Result 3.
  4. [Section IV.C, early-stop threshold] The nine-tool-call threshold for aborting a subgoal is described as 'derived from human performance statistics,' but no derivation or sensitivity analysis is provided. This ad-hoc parameter affects both success and task-completion rates, particularly in long-horizon tasks where the VLM uses 33–43 tool calls. A different threshold could materially change the reported comparisons. The authors should either present the underlying human statistics, or report results for a range of thresholds, to demonstrate that the main conclusions are robust to this choice.
minor comments (4)
  1. [Section V.B] The text says the zero-shot agent 'experiences a drop of approximately 12% in success rate and 37% in completion rate due to map noise.' From Tables I, V, and VI, the absolute success drops are 51.64% → 46.43% (Table V) and 51.64% → 39.29% (Table VI); completion drops are 86.57% → 68.07% and 86.57% → 60.11%. Please specify whether these are percentage-point or relative changes, and correct the numbers accordingly.
  2. [Tables I–IV] The number of VLM trials (187) differs from the number of human trials (198) even though the text says the same 198 tasks were executed by both groups. Clarify why some VLM trials were excluded (e.g., API failures, aborted runs) and whether the difference affects the averaged comparisons.
  3. [Section IV.C] The tool descriptions and full prompt templates are referenced to supplementary material but are not included in the paper. Since the study is about zero-shot VLM behavior, the exact prompts, tool schemas, and model version (GPT-4.1 with tool-calling) are important for reproducibility. Please include them in an appendix or provide a stable link.
  4. [Abstract] The abstract states the system 'offers a deployable framework,' but the evaluation is entirely in simulation and the limitations section lists latency and sparse observation issues as barriers to real-time operation. Please soften 'deployable' to 'a framework for simulated deployment' or support the claim with real-robot results.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical benchmark evaluation, and its metrics and stopping rules are not fitted predictors of the claims they support.

full rationale

This paper reports measurements from a VLM agent and human operators in simulation; it does not derive its headline numbers from a model whose parameters are fit to the same data. Success rate, task completion rate, and SPL are computed directly from episode outcomes, not predicted from a fitted quantity. The use of ground-truth object locations and human experts stopping episodes are experimental supports (oracle evaluation), not inputs that force the VLM's success rate by construction; the VLM still must execute a correct sequence of tools to reach the point where a human declares success. The SPL metric defines l_i as human-expert path length, which makes it a human-normalized efficiency metric, but this is an explicitly stated evaluation choice rather than a hidden reuse of the quantity being claimed. The early-stop threshold of nine tool calls is described as 'derived from human performance statistics,' but it is a stopping rule to avoid infinite loops, not an estimator of VLM success; it does not by itself equalize VLM and human performance, and it is not presented as a prediction. Self-citations such as [3] provide background context and are not load-bearing for the paper's central empirical claims. No self-referential derivation, fitted-then-repredicted quantity, or imported uniqueness theorem appears. The main qualifications (long-horizon collapse, noisy-map degradation) are also measured, not assumed. The evaluation-protocol concerns raised by the reader's take are validity/correctness issues, not circular reasoning, so the circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The main results rest on an oracle semantic map, human supervision, simulation realism, and a human-derived stopping rule; the noisy-map ablations show how much these assumptions matter. The framework introduces no new physical or theoretical entities.

free parameters (1)
  • Subgoal early-stop threshold (nine tool calls) = 9 tool calls
    Set from human performance statistics (Section IV.C, Stopping condition); allows the VLM to abort a stuck subgoal and continue, directly raising reported success and completion rates.
axioms (5)
  • domain assumption Main success numbers assume a semantic map containing ground-truth object locations given to the VLM.
    Section IV.C states ground-truth object locations are provided 'to analyze the VLM's decision-making performance under ideal conditions'; Tables I–IV therefore measure planning with an oracle map, not full autonomy.
  • domain assumption Human experts supervise every episode and manually stop execution when the task is complete.
    Section IV.C, Stopping condition: 'human experts supervise all tasks and manually stop execution when the task is successfully completed.' Success detection is not autonomous.
  • domain assumption Gazebo greenhouse scenes stand in for real horticultural environments.
    Section IV.A describes Gazebo scenes; Section VI concedes the visual appearance differs from real training images, so VLM detector/planner behavior may not transfer.
  • ad hoc to paper Task definition treats a missing target attribute (e.g., a yellow tomato absent on a selected plant) as a failed subgoal.
    Section V.A, Result 1 acknowledges these are 'failures that were artifacts of the task definition rather than actual performance errors', yet they are counted in success rates.
  • domain assumption GPT-4.1's tool-calling and visual reasoning capabilities are taken as given.
    The planner's competence is inherited from a proprietary API; the paper does not ablate over model choice or version.

pith-pipeline@v1.3.0-alltime-deepseek · 16788 in / 12774 out tokens · 127306 ms · 2026-08-03T09:55:10.636341+00:00 · methodology

0 comments
read the original abstract

Crop monitoring is essential for precision agriculture, but current systems lack high-level reasoning. We introduce a novel, modular framework that uses a Vision Language Model (VLM) to guide robotic task planning by actively querying heterogeneous data sources, including enriched RGB camera feeds and 2D semantic occupancy maps, interleaved with robotic action primitives. We contribute a comprehensive benchmark for short- and long-horizon crop monitoring tasks in monoculture and polyculture environments. Our results show that while zero-shot VLMs perform robustly for short-horizon tasks (achieving 87% success, comparable to human experts), success drops significantly to under 10% for complex long-horizon, multi-target tasks. Despite this decline, task completion rates remain above 76% under noiseless conditions. Critically, the system degrades when relying on noisy semantic maps, demonstrating a key limitation in current VLM context grounding for sustained robotic operations. This work offers a deployable framework and critical insights into VLM capabilities and shortcomings for complex agricultural robotics.

Figures

Figures reproduced from arXiv: 2601.11906 by Aditya Potnis, Girish Chowdhary, Jose Cuaran, Kendall Koe, Naveen Kumar Uppalapati.

Figure 1
Figure 1. Figure 1: Sample episode of a single-plant, single-target monitoring task executed by the VLM agent. (a) The VLM takes as input the task prompt, the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall System Pipeline Given context and the assigned task, the agricultural agent selects from a library of action primitives to achieve the task. The agent can switch between manipulation and navigation modes. A library of primitive tasks are available for the agent to choose from to accomplish the task. Once the task is complete or the agent requires help, a human can be prompted to intervene [PITH_FU… view at source ↗
Figure 3
Figure 3. Figure 3: Semantic mapping pipeline. The RGBD data is used to generate an occupancy map of the environment. Detic is used for open vocabulary object [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation Environments Three gazebo simulation environments are used to evaluate the system. (a) A monoculture environment filled with various ripe and unripe tomatoes. (b) A polyculture environment filled with tomatoes, orange peppers, red peppers, and eggplants. (c) A polyculture environments filled with tomatoes, green peppers, red peppers, and eggplants as well as lettuce and strawberries on raised ta… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 19 linked inside Pith

  1. [1]

    Application of precision agri- culture technologies for sustainable crop production and environmental sustainability: A systematic review,

    S. Getahun, H. Kefale, and Y . Gelaye, “Application of precision agri- culture technologies for sustainable crop production and environmental sustainability: A systematic review,”The Scientific World Journal, vol. 2024, no. 1, p. 2126734, 2024

    2024

  2. [2]

    Agricultural robotics: A technical review addressing chal- lenges in sustainable crop production,

    M. Spagnuolo, G. Todde, M. Caria, N. Furnitto, G. Schillaci, and S. Failla, “Agricultural robotics: A technical review addressing chal- lenges in sustainable crop production,”Robotics, vol. 14, no. 2, p. 9, 2025

    2025

  3. [3]

    Active semantic mapping with mobile manipulator in horti- cultural environments,

    J. Cuaran, K. S. Ahluwalia, K. Koe, N. K. Uppalapati, and G. Chowd- hary, “Active semantic mapping with mobile manipulator in horti- cultural environments,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 12 716–12 722

    2025

  4. [4]

    Panoptic mapping with fruit completion and pose estimation for horticultural robots,

    Y . Pan, F. Magistri, T. L¨abe, E. Marks, C. Smitt, C. McCool, J. Behley, and C. Stachniss, “Panoptic mapping with fruit completion and pose estimation for horticultural robots,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 4226–4233

    2023

  5. [5]

    Autonomous apple fruitlet sizing with next best view planning,

    H. Freeman and G. Kantor, “Autonomous apple fruitlet sizing with next best view planning,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 15 847–15 853

    2024

  6. [6]

    3d move to see: Multi-perspective visual servoing towards the next best view within unstructured and occluded environments,

    C. Lehnert, D. Tsai, A. Eriksson, and C. McCool, “3d move to see: Multi-perspective visual servoing towards the next best view within unstructured and occluded environments,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 3890–3897

    2019

  7. [7]

    Viewpoint planning for fruit size and position estimation,

    T. Zaenker, C. Smitt, C. McCool, and M. Bennewitz, “Viewpoint planning for fruit size and position estimation,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 3271–3277

    2021

  8. [8]

    Language models as zero-shot planners: Extracting actionable knowledge for embodied agents,

    W. Huang, P. Abbeel, D. Pathak, and I. Mordatch, “Language models as zero-shot planners: Extracting actionable knowledge for embodied agents,” inInternational conference on machine learning. PMLR, 2022, pp. 9118–9147

    2022

  9. [9]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausmanet al., “Do as i can, not as i say: Grounding language in robotic affordances,”arXiv preprint arXiv:2204.01691, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    Inner monologue: Embodied reasoning through planning with language models,

    W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P. Florence, A. Zeng, J. Tompson, I. Mordatch, Y . Chebotaret al., “Inner monologue: Embodied reasoning through planning with language models,”arXiv preprint arXiv:2207.05608, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,”arXiv preprint arXiv:2209.07753, 2022

    Pith/arXiv arXiv 2022

  12. [12]

    End-to-end navigation with vlms: Transforming spatial reasoning into question-answering,

    D. Goetting, H. G. Singh, and A. Loquercio, “End-to-end navigation with vlms: Transforming spatial reasoning into question-answering,” inWorkshop on Language and Robot Learning: Language as an Interface, 2024

    2024

  13. [13]

    Pivot: Iterative visual prompting elicits actionable knowledge for vlms,

    S. Nasiriany, F. Xia, W. Yu, T. Xiao, J. Liang, I. Dasgupta, A. Xie, D. Driess, A. Wahid, Z. Xuet al., “Pivot: Iterative visual prompting elicits actionable knowledge for vlms,”arXiv preprint arXiv:2402.07872, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    V oronav: V oronoi-based zero-shot object navigation with large language model,

    P. Wu, Y . Mu, B. Wu, Y . Hou, J. Ma, S. Zhang, and C. Liu, “V oronav: V oronoi-based zero-shot object navigation with large language model,” arXiv preprint arXiv:2401.02695, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Closed-loop open-vocabulary mobile manipulation with gpt-4v,

    P. Zhi, Z. Zhang, Y . Zhao, M. Han, Z. Zhang, Z. Li, Z. Jiao, B. Jia, and S. Huang, “Closed-loop open-vocabulary mobile manipulation with gpt-4v,”arXiv preprint arXiv:2404.10220, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Roboexp: Action-conditioned scene graph via interactive ex- ploration for robotic manipulation,

    H. Jiang, B. Huang, R. Wu, Z. Li, S. Garg, H. Nayyeri, S. Wang, and Y . Li, “Roboexp: Action-conditioned scene graph via interactive ex- ploration for robotic manipulation,”arXiv preprint arXiv:2402.15487, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Robot navigation using phys- ically grounded vision-language models in outdoor environments,

    M. Elnoor, K. Weerakoon, G. Seneviratne, R. Xian, T. Guan, M. K. M. Jaffar, V . Rajagopal, and D. Manocha, “Robot navigation using phys- ically grounded vision-language models in outdoor environments,” arXiv preprint arXiv:2409.20445, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Visual language maps for robot navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual language maps for robot navigation,”arXiv preprint arXiv:2210.05714, 2022

    Pith/arXiv arXiv 2022

  19. [19]

    Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action,

    D. Shah, B. Osi ´nski, S. Levineet al., “Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action,” in Conference on robot learning. PMLR, 2023, pp. 492–504

    2023

  20. [20]

    Hierarchical open-vocabulary 3d scene graphs for language-grounded robot navigation,

    A. Werby, C. Huang, M. B ¨uchner, A. Valada, and W. Burgard, “Hierarchical open-vocabulary 3d scene graphs for language-grounded robot navigation,” inFirst Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024

    2024

  21. [21]

    Mapgpt: Map-guided prompting with adaptive path planning for vision-and-language navigation,

    J. Chen, B. Lin, R. Xu, Z. Chai, X. Liang, and K.-Y . K. Wong, “Mapgpt: Map-guided prompting with adaptive path planning for vision-and-language navigation,”arXiv preprint arXiv:2401.07314, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Topv-nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation,

    L. Zhong, C. Gao, Z. Ding, Y . Liao, H. Ma, S. Zhang, X. Zhou, and S. Liu, “Topv-nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation,”arXiv preprint arXiv:2411.16425, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Multimodal language models in agriculture: A tutorial and survey,

    M. Haghighat, A. Saleh, and M. R. Azghadi, “Multimodal language models in agriculture: A tutorial and survey,”Authorea Preprints, 2025

    2025

  24. [24]

    Leveraging llms for mission planning in precision agriculture,

    M. A. Zuzu ´arregui and S. Carpin, “Leveraging llms for mission planning in precision agriculture,”arXiv preprint arXiv:2506.10093, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Strips: A new approach to the appli- cation of theorem proving to problem solving,

    R. E. Fikes and N. J. Nilsson, “Strips: A new approach to the appli- cation of theorem proving to problem solving,”Artificial intelligence, vol. 2, no. 3-4, pp. 189–208, 1971

    1971

  26. [26]

    Hierarchical task and motion planning in the now,

    L. P. Kaelbling and T. Lozano-P ´erez, “Hierarchical task and motion planning in the now,” in2011 IEEE international conference on robotics and automation. IEEE, 2011, pp. 1470–1477

    2011

  27. [27]

    Combined task and motion planning for mobile manipulation,

    J. Wolfe, B. Marthi, and S. Russell, “Combined task and motion planning for mobile manipulation,” inProceedings of the International Conference on Automated Planning and Scheduling, vol. 20, 2010, pp. 254–257

    2010

  28. [28]

    Task and motion planning methods: applications and limitations,

    K. Zhang, E. Lucet, J. A. D. Sandretto, S. Kchir, and D. Filliat, “Task and motion planning methods: applications and limitations,” inICINCO 2022-19th International Conference on Informatics in Control, Automation and Robotics. SCITEPRESS-Science and Tech- nology Publications, 2022, pp. 476–483

    2022

  29. [29]

    Cape: Corrective actions from precondition errors using large language models,

    S. S. Raman, V . Cohen, I. Idrees, E. Rosen, R. Mooney, S. Tellex, and D. Paulius, “Cape: Corrective actions from precondition errors using large language models,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 14 070–14 077

    2024

  30. [30]

    Au- tomatic behavior tree expansion with llms for robotic manipulation,

    J. Styrud, M. Iovino, M. Norrl ¨of, M. Bj ¨orkman, and C. Smith, “Au- tomatic behavior tree expansion with llms for robotic manipulation,” arXiv preprint arXiv:2409.13356, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Openeqa: Embodied question answering in the era of foundation models,

    A. Majumdar, A. Ajay, X. Zhang, P. Putta, S. Yenamandra, M. Henaff, S. Silwal, P. Mcvay, O. Maksymets, S. Arnaudet al., “Openeqa: Embodied question answering in the era of foundation models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16 488–16 498

    2024

  32. [32]

    Enter the mind palace: Reasoning and planning for long-term active embodied question answering,

    M. F. Ginting, D.-K. Kim, X. Meng, A. Reinke, B. J. Krishna, N. Kay- hani, O. Peltzer, D. D. Fan, A. Shaban, S.-K. Kimet al., “Enter the mind palace: Reasoning and planning for long-term active embodied question answering,”arXiv preprint arXiv:2507.12846, 2025

    arXiv 2025

  33. [33]

    Experience is the best teacher: Grounding vlms for robotics through self-generated memory,

    G. Lan, K. Qu, R. Zurbr ¨ugg, C. Chen, C. E. Mower, H. Bou- Ammar, and M. Hutter, “Experience is the best teacher: Grounding vlms for robotics through self-generated memory,”arXiv preprint arXiv:2507.16713, 2025

    arXiv 2025

  34. [34]

    Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning,

    Y . Hu, F. Lin, T. Zhang, L. Yi, and Y . Gao, “Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning,” arXiv preprint arXiv:2311.17842, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Agriclip: Adapting clip for agriculture and live- stock via domain-specialized cross-model alignment,

    U. Nawaz, M. Awais, H. Gani, M. Naseer, F. Khan, S. Khan, and R. M. Anwer, “Agriclip: Adapting clip for agriculture and live- stock via domain-specialized cross-model alignment,”arXiv preprint arXiv:2410.01407, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Agrogpt: Efficient agricultural vision-language model with expert tuning,

    M. Awais, A. H. S. A. Alharthi, A. Kumar, H. Cholakkal, and R. M. Anwer, “Agrogpt: Efficient agricultural vision-language model with expert tuning,” in2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2025, pp. 5687–5696

    2025

  37. [37]

    Visual large language model for wheat disease diagnosis in the wild,

    K. Zhang, L. Ma, B. Cui, X. Li, B. Zhang, and N. Xie, “Visual large language model for wheat disease diagnosis in the wild,”Computers and Electronics in Agriculture, vol. 227, p. 109587, 2024

    2024

  38. [38]

    A vision-language foundation model for leaf disease identification,

    K. N. Quoc, L. L. T. Thu, and L.-D. Quach, “A vision-language foundation model for leaf disease identification,”arXiv preprint arXiv:2505.07019, 2025

    arXiv 2025

  39. [39]

    Few-shot image classification of crop diseases based on vision–language models,

    Y . Zhou, H. Yan, K. Ding, T. Cai, and Y . Zhang, “Few-shot image classification of crop diseases based on vision–language models,” Sensors, vol. 24, no. 18, p. 6109, 2024

    2024

  40. [40]

    Agrobench: Vision-language model benchmark in agriculture,

    R. Shinoda, N. Inoue, H. Kataoka, M. Onishi, and Y . Ushiku, “Agrobench: Vision-language model benchmark in agriculture,”arXiv preprint arXiv:2507.20519, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    A framework for agricultural intelligent analysis based on a visual language large model,

    P. Yu and B. Lin, “A framework for agricultural intelligent analysis based on a visual language large model,”Applied Sciences, vol. 14, no. 18, p. 8350, 2024

    2024

  42. [42]

    Octomap: An efficient probabilistic 3d mapping framework based on octrees,

    A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, and W. Bur- gard, “Octomap: An efficient probabilistic 3d mapping framework based on octrees,”Autonomous robots, vol. 34, pp. 189–206, 2013

    2013

  43. [43]

    High tunnel tomatoes,

    C. Kaiser and M. Ernst, “High tunnel tomatoes,”University of Kentucky College of Agriculture, Food and Environment Cooperative Extension Service, Lexington, 2012

    2012

  44. [44]

    Yields and economics of high tunnels for production of warm-season vegetable crops,

    D. Waterer, “Yields and economics of high tunnels for production of warm-season vegetable crops,”HortTechnology, vol. 13, no. 2, pp. 339–343, 2003

    2003

  45. [45]

    Objectnav revisited: On evaluation of embodied agents navigating to objects,

    D. Batra, A. Gokaslan, A. Kembhavi, O. Maksymets, R. Mottaghi, M. Savva, A. Toshev, and E. Wijmans, “Objectnav revisited: On evaluation of embodied agents navigating to objects,”arXiv preprint arXiv:2006.13171, 2020

    Pith/arXiv arXiv 2006

  46. [46]

    Critical attention scal- ing in long-context transformers,

    S. Chen, Z. Lin, Y . Polyanskiy, and P. Rigollet, “Critical attention scal- ing in long-context transformers,”arXiv preprint arXiv:2510.05554, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models,

    T. Groot and M. Valdenegro-Toro, “Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models,” arXiv preprint arXiv:2405.02917, 2024

    Pith/arXiv arXiv 2024