Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Text-guided Generation of Efficient Personalized Inspection Plans

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A vision-language model can convert a text instruction and a known 3D map into a smooth, collision-free inspection trajectory that visits the described points in the described order.

desk verdict A coherent VLM-guided inspection-planning pipeline with a genuine integration novelty, but the evaluation is too thin to support the strong claims. read the letter →

arxiv 2506.02917 v1 pith:KEYLDQLL submitted 2025-06-03 cs.RO

Xingpeng Sun , Zherong Pan , Xifeng Gao , Kui Wu , Aniket Bera This is my paper
classification cs.RO
keywords inspectionplanningvision-languagemodeltext-guidednavigationtravelingsalesmanproblemtrajectoryoptimizationspatialreasoningprobabilisticroadmapsaliencyfiltering
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

The paper tries to establish a practical equivalence: for known environments, a user's plain-text description of an inspection task can replace the usual hand-authored list of points of interest, viewpoint constraints, and visit order. The proposed pipeline takes a 3D map and a sentence such as "fly around the arena, then go inside to see the soccer pitch" and outputs a smooth, collision-free trajectory for an RGB-equipped aerial or underwater vehicle. A vision-language model does the semantic work, extracting the places mentioned, judging from six rendered views whether a candidate viewpoint actually sees each place with the right spatial relation, and checking smoothed waypoints during refinement, while a Traveling Salesman formulation enforces the desired order and a spline optimizer makes the path executable. The reported experiments show that in the tested scenes the resulting trajectories have lower curvature, fewer segments, and shorter length than the baselines, with comparable or better text-conformity scores. If the claim holds, it would make inspection planning usable by non-experts and adaptable to new scenes without retraining.

What carries the argument

The load-bearing mechanism is the binary predicate $I_{\mathrm{salient}}(n_i, v)$, which decides whether a probabilistic-roadmap node $v$ is an admissible viewpoint for point of interest $n_i$. It is true only when the segmentation mask of $n_i$ in one of the six axis-aligned images taken from $v$ has $\mathrm{IoU} > 0.5$ and the vision-language model confirms both that $n_i$ is visible and that the viewpoint satisfies the prompt's spatial relation (inside, over, in-front, around, or arbitrary). This predicate converts each text-described point into a concrete set $V(n_i)$ of usable waypoints; those waypoint sets are the input to an order-constrained traveling-salesman formulation solved as a mixed-integer program via the Miller-Tucker-Zemlin subtour-elimination method. During smoothing, the same VLM acts as a constraint checker in a backtracking line search, so every trajectory refinement is certified against the original text before the final minimum-snap spline optimization runs.

What would settle it

Render the six axis-aligned views from every roadmap node in a scene, compute ground-truth visibility and spatial-relation labels from the 3D geometry (target inside the camera frustum, camera center inside or outside the POI's bounding volume, and so on), and compare them with the VLM's answers on a set of prompts broader than the paper's four spatial relations; a disagreement rate well above the reported roughly ten percent would mean the waypoint sets feeding the TSP are wrong, and any generated trajectory would visit incorrect or missing locations.

Watch

Extended reading notes

Core claim

The central claim is that a training-free vision-language model can serve as the semantic core of an inspection planner: it converts free-form text into a concrete set of points of interest, identifies sampled viewpoints from which each point is visible and salient, and verifies the user-specified spatial relation during trajectory refinement. For each node of a Poisson-disk-sampled probabilistic roadmap the method captures six axis-aligned images and asks the VLM whether the target appears and whether the viewpoint is inside, over, in front of, around, or arbitrarily placed with respect to it; a node is usable only if the VLM agrees and an open-vocabulary segmentation model reports $\mathrm{IoU} > 0.5$. The paper reports that in six test environments this procedure gives a winning rate above 90% for the four spatial relations with the tested vision-language model and yields trajectories with curvature, step count, and distance clearly better than the PRM+VLM and zero-shot navigation baselines while preserving CLIP-measured text alignment. In the paper's own framing, the contribution is the first training-free VLM-guided inspection planning algorithm for known scenes, producing trajectories that respect user instructions rather than merely reaching a goal.

Load-bearing premise

The load-bearing premise is that a vision-language model, shown only six axis-aligned images with no camera pose or 3D geometry, can reliably judge whether a described location is visible from a candidate viewpoint and whether the viewpoint satisfies the user's spatial phrase; the paper itself admits that its hallucinations can affect reliability.

Editorial extensions

If this is right

  • Inspection tasks can be specified in natural language with no manual waypoints, pose constraints, or ordering, so non-expert users can author inspection plans for known scenes.
  • Because no training or fine-tuning is involved, the same pipeline transfers to new environments and new object vocabularies at zero marginal training cost, bounded only by the vision-language model's competence.
  • The order-constrained TSP formulation guarantees that any feasible route respects the visitation order implied by the text while minimizing travel distance among the chosen waypoints.
  • Re-aiming the camera at the segmentation mask's mass center during execution raises the CLIP text-conformity score by about 0.02 without harming trajectory smoothness.
  • The method applies to aerial and underwater platforms alike, since the output is a continuous collision-free spline with interpolated camera orientations.

Reading between the lines

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

  • Beyond the paper: a stress test the paper does not run is to give the pipeline an impossible or misleading prompt, such as "fly over the sea" inside an indoor map, and measure how the per-node filter degrades; graceful failure in that case remains unexamined.
  • Beyond the paper: the method's reliability is ultimately a property of the VLM's spatial reasoning, so substituting a cheaper or open-source VLM would probably lower the reported spatial-relation accuracy and could serve as a controlled test of how much of the pipeline's success comes from the VLM versus the TSP and optimization stages.
  • Beyond the paper: the pipeline could be inverted into an explainability tool, letting the VLM's per-node reasoning generate a textual explanation of why each viewpoint was chosen.
  • Beyond the paper: the reported hundreds of VLM image queries per scene suggest that API cost, not robot time, is the practical bottleneck; distilling the VLM's viewpoint judgments into a lightweight local classifier is a testable route to real-time deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents a three-stage pipeline for generating inspection trajectories from free-form text and a known 3D map. Stage I uses GPT-4o with multi-view images and GroundedSAM to extract POIs and define valid waypoint sets V(n_i) for each POI under user-specified spatial relations and a saliency threshold. Stage II solves an order-constrained TSP (MTZ formulation) over the PRM graph and applies a VLM-checked iterative smoothing procedure. Stage III computes a minimum-snap collision-free spline trajectory and re-aims the camera using the detected masks. The evaluation on six handcrafted/real-scan environments reports spatial-relation classification accuracy, smoothness/efficiency metrics against PRM+VLM and LM-Nav, CLIP text-conformity scores, and ablations.

Significance. If the claims hold, the work would be a valuable zero-training integration of open-vocabulary perception and reasoning into known-map inspection planning. The pipeline is clearly specified, the use of standard PRM/MTZ-TSP/minimum-snap subroutines is sound, and the evaluation metrics (CLIP, manual spatial labels) are external rather than fitted to the method. The main risk is that the load-bearing per-node predicate Isalient is validated only indirectly, and the trajectory comparisons have no statistical grounding. The work is best viewed as a promising system demonstration whose central empirical claim needs stronger evidence before it can be accepted.

major comments (3)
  1. [III-A, Table I, V] The correctness of the entire pipeline rests on the per-node predicate Isalient(n_i, v) defined in Section III-A, which requires GroundedSAM IoU > 0.5 and VLM confirmation of both visibility and spatial relation. Table I evaluates only the spatial-relation component of this predicate; it does not report precision or recall of the full waypoint sets V(n_i) against geometric ground truth. The MIP constraints (2) in Section III-B make V(n_i) hard constraints, so a false negative removes all valid options for a POI and a false positive injects an unusable waypoint; Section V explicitly concedes VLM hallucinations can affect reliability. I recommend adding an end-to-end evaluation of V(n_i), including per-POI detection rate, precision/recall against manually or geometrically labeled visible nodes, and a sensitivity analysis of the IoU and VLM thresholds.
  2. [IV-B, Tables II and III] The quantitative comparison in Table II and the ablation in Table III are reported as single averages over six simulated environments, without error bars, number of independent runs, per-scene breakdowns, or significance tests. The CLIP scores are close (e.g., 0.2719 versus 0.2897 for PRM+VLM versus Ours), and the claim that smoothing preserves text conformity rests on a difference that is not statistically supported. I request per-scene tables, repeated runs or confidence intervals, and a clarification of how LM-Nav is applied to known maps; a real-robot or at least higher-fidelity simulation check would also make the 'executable' claim more credible.
  3. [III-C] The minimum-snap optimization in Section III-C moves the control points with only a soft penalty to keep them near the original PRM nodes, and the method then re-extracts segmentation masks with GroundedSAM to re-aim the camera. It does not re-verify with the VLM that the final nodes still satisfy Isalient(n_i, v), i.e., that the POI remains visible, salient, and in the required spatial relation. Since the TSP and smoothing stages preserve these properties only on the PRM-restricted path, the final trajectory can in principle violate the text constraints. I recommend adding a post-optimization validity check or hard constraints that preserve V(n_i).
minor comments (4)
  1. [III-B] The constraint numbering is confusing: after constraints (3) and (4), the text refers to the 'forth and fifth constraints' while the displayed equations are (5) and (6); please fix the ordinal and the phrase 'objects stored in node' in the MTZ interpretation.
  2. [III-B] The notation Lo denotes both the ordered list and its cardinality; use |Lo| for the length to avoid ambiguity in constraints (2) and (5).
  3. [IV-C] The CLIP score is taken as the maximum over all camera views, which can mask a POI that is never actually visible; please report per-POI CLIP scores or a visibility-based recall alongside the max.
  4. [Table I] Table I lacks the 'arbitrary' spatial relation and does not report per-environment accuracy; since the six environments are heterogeneous, a per-environment breakdown would help interpret the 90-96% averages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline's stages are independent and evaluated against external CLIP and manual labels.

full rationale

The derivation chain is self-contained rather than circular. Section III-A builds V(n_i) as the set of PRM nodes passing a thresholded GroundedSAM IoU test and a VLM visibility/spatial-relation confirmation; this is a design-defined filtering step, not a fitted parameter later renamed as a prediction. The TSP/MIP in Section III-B and the trajectory optimization in Section III-C minimize geometric objectives (distance, snap) and are evaluated by geometric smoothness metrics, steps, and distance in Table II; these metrics are not fitted to the reported trajectories. Text conformity is measured with CLIP, an independently pretrained model, and spatial-reasoning accuracy in Table I is measured against manually labeled ground truth. The only self-citations, such as [24] (a trajectory-optimization solver co-authored by current authors), are used as an independent algorithmic tool rather than as a load-bearing uniqueness theorem or as justification of the central claim. The paper's own limitation statement about VLM hallucinations is a reliability caveat, not a circular reduction of the method's outputs to its inputs.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The central claim rests on a small set of hand-chosen hyperparameters (sample count, spacing, IoU threshold, alpha, view count) and on three untested domain assumptions: the rendered map represents reality, the VLM's per-node spatial judgments are reliable, and text decomposes into the predefined relation vocabulary. These assumptions are not validated by real-world experiments, so the empirical contribution is narrower than the strong 'training-free' framing suggests.

free parameters (7)
  • PRM sample count = 1000 nodes
    Chosen in Section IV; density controls coverage of possible viewpoints and trajectory cost.
  • Poisson disk minimum node spacing = 10% of bounding box size
    Section III-A; sparsification threshold trades VLM query count against path resolution.
  • Saliency IoU threshold = 0.5
    Section III-A, Isalient definition; a PRM node is valid only if a GroundedSAM mask has IoU > 0.5 in some view.
  • Line-search termination step = alpha_min = 1/8
    Section III-B; stopping threshold for VLM-guided waypoint smoothing.
  • Number of axis-aligned views per node = 6 (±X, ±Y, ±Z)
    Section III-A; the chosen view discretization is load-bearing for visibility judgments.
  • In-context example counts = 1 for POI extraction, 8 for spatial relations
    Section IV; prompt design hyperparameters affecting VLM accuracy.
  • Spatial relation vocabulary = {inside, over, in-front, around, arbitrary}
    Section III-A; limits the set of user requests the system can represent.
assumptions (7)
  • domain assumption The known 3D map M and its rendered images faithfully represent scene geometry and appearance.
    Invoked in Section III-A, where photos are rendered from the mesh or point cloud as proxies for real camera views; no real-world validation is performed (Section V).
  • domain assumption A text instruction can be decomposed into a set of POIs plus optional pose and order constraints within a predefined relation vocabulary.
    Section III-A; this is the assumed input structure. User requests outside this structure are not handled.
  • domain assumption A VLM can infer visibility and spatial relations from six axis-aligned images without 3D geometry or camera pose.
    Section III-A and Figure 2c; the whole waypoint-selection step depends on this empirical premise.
  • domain assumption GroundedSAM IoU is a valid proxy for POI saliency.
    Section III-A; Isalient requires IoU > 0.5, but no independent saliency benchmark is given.
  • standard math The Miller-Tucker-Zemlin MIP formulation correctly enforces connectivity and order constraints.
    Section III-B, Equations (1)-(6); standard textbook result used as a subroutine.
  • standard math The minimum-snap trajectory optimization of Ni et al. [24] yields collision-free, executable paths when warm-started by the PRM nodes.
    Section III-C; adopted from [24] without re-derivation.
  • domain assumption VLM yes/no responses during smoothing preserve text conformity and spatial constraints after node perturbation.
    Section III-B, Stage II; the smoothing loop treats the VLM as the sole constraint checker.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-guided Generation of Efficient Personalized Inspection Plans." pith.science (2026). https://pith.science/paper/KEYLDQLL

@misc{pith2026250602917,
  author       = {Pith},
  title        = {Pith review of: Text-guided Generation of Efficient Personalized Inspection Plans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KEYLDQLL}},
  note         = {Machine review of arXiv:2506.02917}
}
read the original abstract

We propose a training-free, Vision-Language Model (VLM)-guided approach for efficiently generating trajectories to facilitate target inspection planning based on text descriptions. Unlike existing Vision-and-Language Navigation (VLN) methods designed for general agents in unknown environments, our approach specifically targets the efficient inspection of known scenes, with widespread applications in fields such as medical, marine, and civil engineering. Leveraging VLMs, our method first extracts points of interest (POIs) from the text description, then identifies a set of waypoints from which POIs are both salient and align with the spatial constraints defined in the prompt. Next, we interact with the VLM to iteratively refine the trajectory, preserving the visibility and prominence of the POIs. Further, we solve a Traveling Salesman Problem (TSP) to find the most efficient visitation order that satisfies the order constraint implied in the text description. Finally, we apply trajectory optimization to generate smooth, executable inspection paths for aerial and underwater vehicles. We have evaluated our method across a series of both handcrafted and real-world scanned environments. The results demonstrate that our approach effectively generates inspection planning trajectories that adhere to user instructions.

Figures

Figures reproduced from arXiv: 2506.02917 by the authors.

Figure 1
Figure 1. Our method generates target inspection planning trajectory (red) that conforms to the text description (top left). The text description contains a set of POIs (dashed yellow boxes), the relative robot poses required to observe these POIs (green), the desired visitation order of POIs (blue), and the bottom images show these locations in the first person. This limitation highlights the need for intuitive, high-level i… view at source ↗
Figure 2
Figure 2. The pipeline of our method. The input to our method is a text description and a 3D environmental map (a). Our method would first interact with VLM to extract a set of POIs L (b). Next, we construct a PRM and query each PRM node v from which the POI ni is visible and salient, while also ensuring that the node satisfies optionally specified robot poses (c). We further ask the VLM to output a pixel-wise segmentation ma… view at source ↗
Figure 3
Figure 3. A heatmap visualization of the VLM’s spatial understanding of the instruction “fly inside the stadium," where red and green indicates areas considered by VLM to be outside and inside the stadium, respectively. Six images are taken at every sample point and forwarded to the VLM. The six images from the two example views are shown on the top and bottom. B. Stage II: PRM-restricted Trajectory Generation We define a PRM… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Outputs of our method, where the inspection trajectories are drawn in red. Robot agent viewpoint camera frames of selected POIs are attached on the left side to highlight text conformity, with the corresponding orientations marked along the trajectory. More visual comp…
Figure 5
Figure 5. Figure 5: Results under different visitation order constraints. To follow personalized user text guidance, the left trajectory visits the “harbor with boat" before the “American flag," while the right trajectory visits the “American flag" first. C. Text Conformity The text confo…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 30 canonical work pages

  1. [1]

    A survey on inspecting structures using robotic systems

    Randa Almadhoun, Tarek Taha, Lakmal Seneviratne, Jorge Dias, and Guowei Cai. A survey on inspecting structures using robotic systems. International Journal of Advanced Robotic Systems , 13(6):1729881416663664, 2016

    work page 2016

  2. [2]

    Ergodic coverage in constrained environments using stochastic trajectory optimization

    Elif Ayvali, Hadi Salman, and Howie Choset. Ergodic coverage in constrained environments using stochastic trajectory optimization. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5204–5210. IEEE, 2017

    work page 2017

  3. [3]

    Robotic tools for deep water archaeology: Surveying an ancient shipwreck with an autonomous underwater vehicle

    Brian Bingham, Brendan Foley, Hanumant Singh, Richard Camilli, Katerina Delaporta, Ryan Eustice, Angelos Mallios, David Mindell, Christopher Roman, and Dimitris Sakellariou. Robotic tools for deep water archaeology: Surveying an ancient shipwreck with an autonomous underwater vehicle. Journal of Field Robotics , 27(6):702–717, 2010

    work page 2010

  4. [4]

    Structural inspection path planning via iterative viewpoint resampling with application to aerial robotics

    Andreas Bircher, Kostas Alexis, Michael Burri, Philipp Oettershagen, Sammy Omari, Thomas Mantel, and Roland Siegwart. Structural inspection path planning via iterative viewpoint resampling with application to aerial robotics. In 2015 IEEE International Conference on Robotics and Automation (ICRA) , pages 6423–6430. IEEE, 2015

    work page 2015

  5. [5]

    Structural inspection path planning via iterative viewpoint resampling with application to aerial robotics

    Andreas Bircher, Kostas Alexis, Michael Burri, Philipp Oettershagen, Sammy Omari, Thomas Mantel, and Roland Siegwart. Structural inspection path planning via iterative viewpoint resampling with application to aerial robotics. In 2015 IEEE International Conference on Robotics and Automation (ICRA) , pages 6423–6430, 2015

    work page 2015

  6. [6]

    A gradient-based inspection path optimization approach

    Boris Bogaerts, Seppe Sels, Steve Vanlanduit, and Rudi Penne. A gradient-based inspection path optimization approach. IEEE Robotics and Automation Letters , 3(3):2646–2653, 2018

    work page 2018

  7. [7]

    Tim Danner and Lydia E. Kavraki. Randomized planning for short inspection paths. Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065) , 2:971–976 vol.2, 2000

    work page 2000

  8. [8]

    Improvements and extensions to the miller-tucker-zemlin subtour elimination constraints

    Martin Desrochers and Gilbert Laporte. Improvements and extensions to the miller-tucker-zemlin subtour elimination constraints. Operations Research Letters, 10(1):27–36, 1991

    work page 1991

Show all 42 references
  1. [9]

    Multiregion inspection by combining clustered traveling salesman tours with sampling-based motion planning

    Stefan Edelkamp, Mihai Pomarlan, and Erion Plaku. Multiregion inspection by combining clustered traveling salesman tours with sampling-based motion planning. IEEE Robotics and Automation Letters, 2(2):428–435, 2017

    2017

  2. [10]

    Computationally- efficient roadmap-based inspection planning via incremental lazy search

    Mengyu Fu, Oren Salzman, and Ron Alterovitz. Computationally- efficient roadmap-based inspection planning via incremental lazy search. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 7449–7456. IEEE, 2021

    2021

  3. [11]

    Asymptotically optimal sampling-based motion planning methods

    Jonathan D Gammell and Marlin P Strub. Asymptotically optimal sampling-based motion planning methods. Annual Review of Control, Robotics, and Autonomous Systems , 4(1):295–318, 2021

    2021

  4. [12]

    Vision-and-language navigation: A survey of tasks, methods, and future directions

    Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Eric Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. arXiv preprint arXiv:2203.12667 , 2022

    2022 arXiv

  5. [13]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 , 2021

    2021 arXiv

  6. [14]

    3d concept learning and reasoning from multi-view images

    Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9202–9212, 2023

    2023

  7. [15]

    Random inspection tree algorithm in visual inspection with a realistic sensing model and differential constraints

    Pˇremysl Kafka, Jan Faigl, and Petr Vá ˇna. Random inspection tree algorithm in visual inspection with a realistic sensing model and differential constraints. In 2016 IEEE International Conference on Robotics and Automation (ICRA) , pages 2782–2787, 2016

    2016

  8. [16]

    Sampling-based algorithms for optimal motion planning

    Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846–894, 2011

    2011

  9. [17]

    Language models as zero-shot trajectory generators

    Teyun Kwon, Norman Di Palo, and Edward Johns. Language models as zero-shot trajectory generators. IEEE Robotics and Automation Letters, 2024

    2024

  10. [18]

    Spherical barycentric coordinates

    Torsten Langer, Alexander Belyaev, and Hans-Peter Seidel. Spherical barycentric coordinates. In Symposium on Geometry Processing , pages 81–88, 2006

    2006

  11. [19]

    Code as policies: Language model programs for embodied control

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 9493–9500. IEEE, 2023

    2023

  12. [20]

    Ok-robot: What really matters in integrating open-knowledge models for robotics

    Peiqi Liu, Yaswanth Orru, Chris Paxton, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Ok-robot: What really matters in integrating open-knowledge models for robotics. arXiv preprint arXiv:2401.12202, 2024

    2024 arXiv

  13. [21]

    Aerialvln: Vision-and-language navigation for uavs

    Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yanning Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15384–15394, 2023

    2023

  14. [22]

    Minimum snap trajectory generation and control for quadrotors

    Daniel Mellinger and Vijay Kumar. Minimum snap trajectory generation and control for quadrotors. In 2011 IEEE international conference on robotics and automation , pages 2520–2525. IEEE, 2011

    2011

  15. [23]

    Simple open-vocabulary object detection with vision transformers

    M Minderer, A Gritsenko, A Stone, M Neumann, D Weissenborn, A Dosovitskiy, A Mahendran, A Arnab, M Dehghani, Z Shen, et al. Simple open-vocabulary object detection with vision transformers. arxiv

  16. [24]

    Robust & asymptotically locally optimal uav-trajectory generation based on spline subdivision

    Ruiqi Ni, Teseo Schneider, Daniele Panozzo, Zherong Pan, and Xifeng Gao. Robust & asymptotically locally optimal uav-trajectory generation based on spline subdivision. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages 7715–7721. IEEE, 2021

    2021

  17. [25]

    Numerical optimization

    Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999

    1999

  18. [26]

    Poisson-rrt

    Chonhyon Park, Jia Pan, and Dinesh Manocha. Poisson-rrt. In 2014 IEEE International Conference on Robotics and Automation (ICRA) , pages 4667–4673, 2014

    2014

  19. [27]

    Survey of uav motion planning

    Lun Quan, Luxin Han, Boyu Zhou, Shaojie Shen, and Fei Gao. Survey of uav motion planning. IET Cyber-systems and Robotics , 2(1):14–21, 2020

    2020

  20. [28]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , p...

    2021

  21. [29]

    Off-line view planning for the inspection of mechanical parts

    Roberto Raffaeli, Maura Mengoni, Michele Germani, and Ferruccio Mandorli. Off-line view planning for the inspection of mechanical parts. International Journal on Interactive Design and Manufacturing (IJIDeM), 7:1–12, 2013

    2013

  22. [30]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 779–788, 2016

    2016

  23. [31]

    Grounded sam: Assembling open-world models for diverse visual tasks

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024

    2024 arXiv

  24. [32]

    Semantic trajectory planning for long-distant unmanned aerial vehicle navigation in urban environments

    Markus Ryll, John Ware, John Carter, and Nick Roy. Semantic trajectory planning for long-distant unmanned aerial vehicle navigation in urban environments. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 1551–1558. IEEE, 2020

    2020

  25. [33]

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

    Dhruv Shah, Bła ˙zej Osi ´nski, Sergey Levine, et al. Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action. In Conference on robot learning , pages 492–504. PMLR, 2023

    2023

  26. [34]

    Fast uav trajectory optimization using bilevel optimization with analytical gradients

    Weidong Sun, Gao Tang, and Kris Hauser. Fast uav trajectory optimization using bilevel optimization with analytical gradients. IEEE Transactions on Robotics , 37(6):2010–2024, 2021

    2010

  27. [35]

    Beyond text: Utilizing vocal cues to improve decision making in llms for robot navigation tasks

    Xingpeng Sun, Haoming Meng, Souradip Chakraborty, Amrit Bedi, and Aniket Bera. Beyond text: Utilizing vocal cues to improve decision making in llms for robot navigation tasks. Transactions on Machine Learning Research

  28. [36]

    Trustnavgpt: Modeling uncertainty to improve trustworthiness of audio-guided llm-based robot navigation

    Xingpeng Sun, Yiran Zhang, Xindi Tang, Amrit Singh Bedi, and Aniket Bera. Trustnavgpt: Modeling uncertainty to improve trustworthiness of audio-guided llm-based robot navigation. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 8794–88...

    2024

  29. [37]

    Holy grail 2.0: From natural language to constraint models

    Dimos Tsouros, Hélène Verhaeghe, Serdar Kadıo ˘glu, and Tias Guns. Holy grail 2.0: From natural language to constraint models. arXiv preprint arXiv:2308.01589, 2023

    2023 arXiv

  30. [38]

    A smart surface inspection system using faster r-cnn in cloud-edge computing environment

    Yuanbin Wang, Minggao Liu, Pai Zheng, Huayong Yang, and Jun Zou. A smart surface inspection system using faster r-cnn in cloud-edge computing environment. Advanced Engineering Informatics, 43:101037, 2020

    2020

  31. [39]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems , 35:24824–24837, 2022

    2022

  32. [40]

    Hallucination is inevitable: An innate limitation of large language models

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024

    2024 arXiv

  33. [41]

    Navgpt: Explicit reasoning in vision-and-language navigation with large language models

    Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 7641–7649, 2024

    2024

  34. [2022]

    arXiv preprint arXiv:2205.06230 , 2, 2022

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.