EdgeAI Drone for Autonomous Construction Site Demonstrator

Arda Taha Candan; Co\c{s}kun An{\i}l Zaman; Emre Girgin

REVIEW 4 major objections 5 minor 1 cited by

EdgeAI Drone for Autonomous Construction Site Demonstrator

T0 review · 4 major / 5 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 custom hexacopter with on-board edge-AI detection and 5G coordination rerouted construction vehicles in a field demonstration.

desk verdict A construction-drone demonstrator with a real integration and a public dataset, whose headline MCU real-time claim is contradicted by its own field-test section. read the letter →

arxiv 2505.09837 v1 pith:XZ3S7ZSI submitted 2025-05-14 cs.RO

Emre Girgin , Arda Taha Candan , Coşkun An{i}l Zaman This is my paper · ORCID

classification cs.RO

keywords edge-AIobjectdetectionUAVsurveillanceconstructionsitemulti-robotcoordinationquantization5Gdataset

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 describes a demonstrator in which a custom-built hexacopter surveys a construction site, detects people and obstacles with an on-board edge-AI detector, and sends their global locations over a 5G link to a central coordinator that reroutes an autonomous excavator and a UGV. The authors want to show that low-power edge hardware, specifically a microcontroller-class device running a quantized object-detection model, can carry the real-time perception load that currently forces drones to rely on heavier computers. They also release a custom dataset of roughly 25,000 training and 5,000 test images designed for small input resolutions, along with a monocular altitude-to-distance calibration method. A field test at a 50-by-100-meter test site demonstrated the full detection-to-rerouting loop over about 250 meters of vehicle travel. The paper also reports that the fastest microcontroller model runs at 0.85 frames per second and states that the field test used the drone's faster general-purpose processor for the AI, so the MCU-only real-time claim is a target rather than a demonstrated result.

What carries the argument

The key machinery is the quantized lightweight detector YoloLC-192, shown in Table I as the fastest model on the microcontroller at 0.85 FPS, paired with a RANSAC-trained polynomial that estimates ground distance from drone altitude and pixel size. This calibration lets a single monocular top-down camera place detected objects in global coordinates using GNSS data, so the central coordinator can treat them as dynamic obstacles. The coordination side rests on MQTT messaging over 5G and a Voronoi-graph A* path planner that recomputes vehicle routes from the drone's obstacle updates. In the field test, the inference step ran on the A72 rather than the microcontroller, so the demonstrated real-time loop depends on that faster processor.

What would settle it

Run the same field scenario with only the microcontroller performing detection while a person walks through the planned vehicle path; if the coordinator cannot reroute the vehicles before the person crosses the safety margin, the MCU-only real-time claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a quantized, small-footprint object-detection model running on a microcontroller can provide real-time obstacle and human-presence detection from a monocular UAV camera, and that this detection, sent through 5G, is enough for a central coordinator to keep construction vehicles collision-free. The authors validate the architecture with a full demonstrator: a hexacopter with a microcontroller-class board, an onboard computer, a 5G router, and a flight controller; an excavator and a UGV executing waypoint commands; and a Ground Control Station that uses Voronoi-graph A* path planning and MQTT messaging over 5G. They also introduce a custom dataset and a RANSAC-trained polynomial that maps image scale to real-world distance from drone altitude, which converts monocular detections into global obstacle coordinates. The paper's own data qualify the claim: on the microcontroller, the best model runs at 0.85 FPS, so the authors explicitly state that field tests used the A72 processor to run the AI. The contribution is therefore best understood as a demonstrated end-to-end coordination system plus a dataset and calibration method, with microcontroller-level real-time detection as the motivating goal the architecture is built around.

Load-bearing premise

The load-bearing premise is that the microcontroller's best detection rate of roughly 0.85 frames per second is fast enough for the path-planning and collision-avoidance loop, because the field test ran detection on a faster onboard processor and never exercised the microcontroller in the closed loop.

Editorial extensions

If this is right

If the microcontroller inference speed reaches real time, the same 5G coordination loop can detect a person and reroute ground vehicles without any GPU or ground-station processing.
The custom dataset, built for small input resolutions, lets other teams train and compare tiny detectors for construction-site surveillance without re-collecting imagery.
The altitude-to-scale calibration gives monocular drones a practical way to produce global obstacle positions, reducing the need for stereo or LiDAR on the aerial platform.
Because ground vehicles speak MQTT over 5G, adding more vehicles or a second drone to the site requires no change to the UAV's onboard perception pipeline.
The field tests point to roughly 8 meters as the operating altitude that balances low-altitude false cone detections against high-altitude misses, which is directly usable in deployment planning.

Reading between the lines

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

A practical near-term system would likely keep the A72-class processor for inference and use the microcontroller for sensor abstraction, because the reported 0.85 FPS on the MCU would make obstacle updates too slow for a moving pedestrian.
The paper's RANSAC scale estimator could be reused outside construction sites, for example in warehouse or disaster-response surveys, but its accuracy at altitudes above 15 meters is untested in this work.
A direct test of the closed loop with MCU-only inference and a person walking through the vehicle path would separate the architecture's value from the hardware's current limits.
The dataset's GNSS metadata, described as useful for visual odometry, could support future work on metric monocular localization for small drones even when altitude is less stable.

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.

Desk Editor's Note

read the letter

The paper is best read as an engineering demonstrator, not a research claim. The headline contribution in the abstract and introduction—MCU-based real-time obstacle detection, validated in a realistic environment—is not supported by the paper's own data. Table I shows the fastest STM32 result at 0.85 FPS, and Section IV-B explicitly says the field tests ran the AI on the A72. That is a load-bearing mismatch. The stress-test note is right about that, and it is the main thing to know about this paper.

That said, there is real work here. The integrated system—custom hexacopter, Pixhawk, STM32 board, 5G/MQTT coordination, Voronoi-A* planner, autonomous excavator and UGV—is a nontrivial build, and the paper reports a 250-meter field operation where the drone's detections rerouted ground vehicles. The RANSAC altitude-to-scale mapping is a sensible idea, and Table II gives RMSE values on unseen altitudes, which is more than many integration papers do. The TUBITAK-EdgeDrone dataset, with roughly 25k training and 5k test images plus GNSS metadata, is publicly available and could be a useful resource for MCU-oriented aerial vision. The paper also honestly admits the MCU limitation in IV-B, which earns some trust.

The soft spots are concentrated around that admission. No detection accuracy (precision/recall) is reported anywhere. There is no end-to-end latency measurement from camera frame to updated waypoint. The field test is qualitative; we do not learn whether the A72-based system actually prevented a collision or just produced a path change. The real-time requirement is never defined, so we cannot tell what threshold the MCU would need to meet. The dataset is introduced but not benchmarked against existing detectors, so its value as an artifact is plausible but unproven.

Proportionately, the central claim fails as stated, but the narrower claim—that an edge-AI drone with a Raspberry Pi-class processor can feed obstacle locations into a 5G-connected multi-robot coordination loop—is demonstrated at a qualitative level. The paper is salvageable if the authors reframe it as exactly that, add detection accuracy and latency numbers, and either exercise the MCU in the loop or remove the MCU real-time claim entirely.

Who is this for? Groups working on low-cost construction-site monitoring, edge AI on UAVs, or multi-robot coordination who want a concrete demonstrator and a possible training dataset. It deserves a serious referee: the integration is real, the artifact is public, and the flaws are fixable. I would send it to review with a request for major revision rather than desk-reject, and I would tell the authors to bring the claims in line with what they actually measured.

Referee Report

4 major / 5 minor

Summary. The paper describes an integrated demonstrator for autonomous multi-robot coordination on a construction site, comprising a custom hexacopter with an onboard STM32 microcontroller, a 5G-connected ground control station, an autonomous excavator, and a custom UGV. The authors propose lightweight, quantized object detection models deployed on the MCU, a monocular RANSAC-based pixel-to-meter scale model for localizing detected objects, and a publicly released dataset (TUBITAK-EdgeDrone). They report frame-rate benchmarks on three processors, RANSAC mapping errors at several altitudes, and a field test in which the drone detects people and the coordinator replans vehicle routes. The central asserted contribution is real-time MCU-based obstacle and human-presence detection validated in a realistic construction environment.

Significance. If the claims were fully supported, the paper would offer a useful data point for low-power edge-AI in construction robotics, a public dataset for small-footprint detection, and a working multi-agent integration on 5G. The public dataset and the full-system field integration are potentially reusable artifacts, and the explicit frame-rate benchmark table across i7, A72, and M7 is a useful negative result for MCU-based real-time detection. However, the paper's headline claim is undercut by its own field-test description, and several quantitative evaluation components are missing. The significance of the work as a research contribution therefore depends on a substantial revision that aligns the claims with the evidence and adds missing metrics.

major comments (4)

[Abstract, Section I, Section IV-B, Table I] The central claim that real-time obstacle and human-presence detection was implemented and validated on constrained MCU hardware is directly contradicted by the field-test description. Section IV-B states: 'Since the performance of the detection models are not enough to satisfy real-time requirements, as shown in the Table I, we used the A72 to run the AI during field tests.' Table I shows the fastest MCU result (YoloLC-192 on the STM32 M7) at 0.85 FPS. Consequently, the closed-loop collision-avoidance and path-replanning demonstration was executed on the Raspberry Pi-class A72, not on the MCU. The abstract and the contribution list in Section I therefore overstate what was validated. The paper must either report an experiment that exercises the MCU inside the closed loop, or substantially reframe the central contribution to the A72-based edge-AI system and present the MCU results as a benchmark of current limitations.
[Section IV-A, Tables I and II] No detection accuracy is reported for any model on any dataset. Table I reports only frame rates, and Table II reports RANSAC RMSE values, which are conditioned on correctly detected cone markers. Without precision, recall, or mAP on the TUBITAK-EdgeDrone test split, the reader cannot judge whether the detector reliably finds humans or obstacles at the operating altitudes used in the field test. Detection failures are especially important here because the RANSAC localization error is meaningless for missed detections. The authors should add standard detection metrics and, ideally, report detection accuracy at the chosen operational altitude of approximately 8 m.
[Section IV-B, Section I] The term 'real-time' is never defined, and no end-to-end latency is measured. The paper does not specify a required frame rate or a maximum acceptable delay between image capture, object detection, localization, transmission to the GCS, and updated waypoint dispatch. Without this definition, the statement that the MCU's 0.85 FPS is 'not enough to satisfy real-time requirements' is an unquantified judgment, and the substitution of the A72 in the field test cannot be evaluated against a concrete requirement. The authors should define a quantitative real-time criterion for the dynamic path-planning loop and report the corresponding end-to-end latency for the deployed configuration.
[Section IV-B, Figure 4] The field-test evaluation is only qualitative. The text describes the scenario and states that the drone relayed detected individuals and the GCS adjusted routes, but it reports no quantitative outcome measures such as the number of successful replans, minimum separation distances, detection-to-replan delays, or the total distance traveled under replanning. The claim that 'field experiments demonstrate practical viability' is therefore supported only by narrative. Quantitative field-test logs, or at least a table of measured event timings and route deviations, should be added.

minor comments (5)

[Section III-D] The text refers to a 'RANSAC model' trained on cone-marked data. This should be clarified as a polynomial regression fitted with RANSAC, with the polynomial degree and the number of inlier/outlier samples specified.
[Table II] Table II reports RMSE values as point estimates without the number of test samples, standard deviations, or confidence intervals. This makes it hard to judge whether the cubic model is significantly better than the quadratic or quartic alternatives.
[Section IV-A and Section III-C] The hardware names are inconsistent: Section III-C mentions an 'STM32F769i-Discovery' board, while Section IV-A refers to an 'STM32H769-I with M7-Cortex.' The same applies to the general-purpose processor being called both 'i7-8700' and 'CPU i7-8700.' The authors should standardize these names and verify the exact STM32 part number.
[Abstract and Section IV-B] There are typographical issues such as 'UA V' with a spurious space in the abstract and elsewhere, and the section heading 'C ONCLUSION ACKNOWLEDGEMENT' should be split into separate 'Conclusion' and 'Acknowledgment' sections.
[Figure 4] The text says the route, loading site, and dumping site are shown in Figure 4, but the figure caption only describes the OpenAirLab area. The route overlay and site labels should be made explicit in the figure or a new figure should be added.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the RANSAC and detection results are validated on unseen data, and the MCU-versus-A72 field-test discrepancy is an evidence gap, not a definitional loop.

full rationale

The paper's derivation chain is not circular. The RANSAC altitude-to-pixel-scale model is fitted on cone-marked site data and then evaluated on unseen data at altitudes 5, 8, 10, and 15 m (Table II), which is an external validation rather than a fitted input renamed as a prediction. Object detection models are trained on the TUBITAK-EdgeDrone dataset and benchmarked across processors (Table I); no model output is defined in terms of the evaluation target. The only self-citation, [12] (Sari et al.), appears as related-work context for construction robots, and the paper does not rely on it for any load-bearing derivation, uniqueness theorem, or ansatz. The discrepancy that real-time field experiments used the A72 rather than the STM32 MCU (Section IV-B) is a gap between the claimed MCU-based real-time contribution and the demonstrated closed-loop evidence; it concerns validity and correctness, not circularity. There is no equation, definition, or fitted parameter that reduces a reported result to its input by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central demonstration rests on standard robotics infrastructure plus a set of fitted or assumed quantities: the RANSAC scale-to-altitude mapping (fitted to cone-marked data), the selected 8 m operating altitude, and the assumption that flat-ground geometry and reliable 5G/GNSS delivery hold. The paper introduces no new theoretical entities; its inventions are hardware and a dataset.

free parameters (2)

RANSAC polynomial coefficients for pixel-to-meter scale vs altitude
Trained on cone-marked site data (Section IV-A), used to convert image pixel sizes to world coordinates; the paper reports only RMSE per degree, not coefficients.
Optimal operational flight altitude = 8 m
Selected post hoc from Table II RMSE results; the paper states 8 meters was chosen for testing (Section IV-A).

assumptions (4)

domain assumption Monocular top-down camera at known altitude with a locally flat ground plane supports pixel-dimension-to-distance mapping.
Section III-D states the task is ill-posed for structure-from-motion and relies on altitude and object pixel dimensions; if the ground is not flat this localization fails.
domain assumption The custom RANSAC relation learned on a cone-marked site generalizes to the OpenAirLab field site.
The field test depends on transferring the learned scale-vs-altitude map without retraining or uncertainty bounds (Sections IV-A and IV-B).
ad hoc to paper Detection performance measured on the A72 processor is representative of the edge-AI system's real-time capability in the field.
This is not stated as an assumption but is required to bridge Table I MCU FPS values (0.85 FPS) with the field-test statement that the A72 ran the AI (Section IV-B).
domain assumption 5G/MQTT communication and GNSS geodesic conversions are sufficiently accurate for collision-free path planning.
The architecture assumes reliable low-latency delivery of obstacle locations to the GCS for Voronoi-A* planning; no latency or positioning error analysis is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EdgeAI Drone for Autonomous Construction Site Demonstrator." pith.science (2026). https://pith.science/paper/XZ3S7ZSI

@misc{pith2026250509837,
  author       = {Pith},
  title        = {Pith review of: EdgeAI Drone for Autonomous Construction Site Demonstrator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZ3S7ZSI}},
  note         = {Machine review of arXiv:2505.09837}
}

read the original abstract

The fields of autonomous systems and robotics are receiving considerable attention in civil applications such as construction, logistics, and firefighting. Nevertheless, the widespread adoption of these technologies is hindered by the necessity for robust processing units to run AI models. Edge-AI solutions offer considerable promise, enabling low-power, cost-effective robotics that can automate civil services, improve safety, and enhance sustainability. This paper presents a novel Edge-AI-enabled drone-based surveillance system for autonomous multi-robot operations at construction sites. Our system integrates a lightweight MCU-based object detection model within a custom-built UAV platform and a 5G-enabled multi-agent coordination infrastructure. We specifically target the real-time obstacle detection and dynamic path planning problem in construction environments, providing a comprehensive dataset specifically created for MCU-based edge applications. Field experiments demonstrate practical viability and identify optimal operational parameters, highlighting our approach's scalability and computational efficiency advantages compared to existing UAV solutions. The present and future roles of autonomous vehicles on construction sites are also discussed, as well as the effectiveness of edge-AI solutions. We share our dataset publicly at github.com/egirgin/storaige-b950

Figures

Figures reproduced from arXiv: 2505.09837 by the authors.

**Figure 2.** The overall architecture of the demonstrator is summarized. While the GCS coordinates all the vehicles, communication is facilitated by a 5G [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The CAD drawing of the hexacopter, including the AI subsystem. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: OpenAirLab is an autonomous construction test site in Gebze, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

UAVs Meet Agentic AI: A Multidomain Survey of Autonomous Aerial Intelligence and Agentic UAVs
cs.RO 2025-06 conditional novelty 3.0 of 10

A narrative survey defines 'agentic UAVs' as drones with perception, cognition, control, and communication layers and catalogs applications and challenges across eight domains.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages · cited by 1 Pith paper

[1]

Mobile Manipulator for Autonomous Localization, Grasping and Precise Placement of Construction Material in a Semi-Structured Environment,

ˇStibinger, Petr and Broughton, George and Majer, Filip and Rozsyp´alek, Zden ˇek and Wang, Anthony and Jindal, Kshitij and Zhou, Alex and Thakur, Dinesh and Loianno, Giuseppe and Krajn ´ık, Tom´aˇs and Saska, Martin, “Mobile Manipulator for Autonomous Localization, Grasping and Precise Placement of Construction Material in a Semi-Structured Environment,”...

work page 2021
[2]

Perceptive Model Predictive Control for Continuous Mobile Manipulation,

Pankert, Johannes and Hutter, Marco, “Perceptive Model Predictive Control for Continuous Mobile Manipulation,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 6177-6184, 2020

work page 2020
[3]

Large-scale 3D printing by a team of mobile robots,

Zhang, Xu and Li, Mingyang and Lim, Jian Hui and Weng, Yiwei and Tay, Yi Wei Daniel and Pham, Hung and Pham, Quang-Cuong, “Large-scale 3D printing by a team of mobile robots,” Automation in Construction, vol. 95, pp. 98–106, 2018

work page 2018
[4]

Tie a knot: human–robot cooperative workflow for assembling wooden structures using rope joints,

Mitterberger, Daniela and Atanasova, Lidia and D ¨orfler, Kathrin and Gramazio, Fabio and Kohler, Matthias, “Tie a knot: human–robot cooperative workflow for assembling wooden structures using rope joints,” Construction Robotics, vol. 6, no. 3, pp. 277–292, 2022

work page 2022
[5]

Accurate and Adaptive in Situ Fabrication of an Undulated Wall Using an on-Board Visual Sensing System,

Lussi, Manuel and Sandy, Timothy and D ¨orfler, Kathrin and Hack, Norman and Gramazio, Fabio and Kohler, Matthias and Buchli, Jonas, “Accurate and Adaptive in Situ Fabrication of an Undulated Wall Using an on-Board Visual Sensing System,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), 2018, pp. 3532-3539

work page 2018
[6]

Robot for automatic waste sorting on construction sites,

Chen, Xinxing and Huang, Huaiyang and Liu, Yuxuan and Li, Jiqing and Liu, Ming, “Robot for automatic waste sorting on construction sites,” Automation in Construction, vol. 141, pp. 104387, 2022

work page 2022
[7]

Intelligent demolition robot: Structural statics, collision detection, and dynamic control,

Mu, Zonggao and Liu, Liyuan and Jia, Lihui and Zhang, Luyang and Ding, Ning and Wang, Chengjiang, “Intelligent demolition robot: Structural statics, collision detection, and dynamic control,” Automa- tion in Construction, vol. 142, pp. 104490, 2022

work page 2022
[8]

Robots on construction sites: The potential and challenges of on-site digital fabrication,

Loveridge, Russell and Coray, Tanja, “Robots on construction sites: The potential and challenges of on-site digital fabrication,” Science Robotics, vol. 2, no. 5, pp. eaan3674, 2017

work page 2017

Show all 19 references

[9]

Artificial intelligence in green building,

Debrah, Caleb and Chan, Albert P.C. and Darko, Amos, “Artificial intelligence in green building,” Automation in Construction, vol. 137, pp. 104192, 2022

2022
[10]

A survey on teaching workplace skills to construction robots,

Wu, Haitao and Li, Heng and Fang, Xin and Luo, Xiaochun, “A survey on teaching workplace skills to construction robots,” Expert Systems with Applications, vol. 205, pp. 117658, 2022

2022
[11]

Recent advancements of robotics in construction,

Xiao, Bo and Chen, Chen and Yin, Xianfei, “Recent advancements of robotics in construction,” Automation in Construction, vol. 144, pp. 104591, 2022

2022
[12]

Cloud assisted connected and automated mobility system architecture design and experimental verification: The 5G-MOBIX autonomous truck routing use case,

Sari, Tahir and Sever, Mert and Candan, Arda Taha and Girgin, Tuba and Girgin, Emre and Haklidir, Mehmet, “Cloud assisted connected and automated mobility system architecture design and experimental verification: The 5G-MOBIX autonomous truck routing use case,” in IEEE Intelli...

2022
[13]

Deep-learning- based vision for earth-moving automation,

Borngrund, Carl and Sandin, Fredrik and Bodin, Ulf, “Deep-learning- based vision for earth-moving automation,” Automation in Construc- tion, vol. 133, pp. 104013, 2022

2022
[14]

2nd Future of Construction Workshop at the Int

Jin, Shiyu and Kang, Liuwang and Peng, Yuan-Chih and He, Zhenpeng and Song, Ruitao and Qian, Lingfeng and Zhang, Liangjun, ”Au- tonomous Excavator System for Construction Earth Moving,” in Proc. 2nd Future of Construction Workshop at the Int. Conf. on Robotics and Automation (...

2023
[15]

Learning excavation of rigid objects with offline reinforcement learning,

Jin, Shiyu and Ye, Zhixian and Zhang, Liangjun, “Learning excavation of rigid objects with offline reinforcement learning,” arXiv preprint arXiv:2303.16427, 2023

2023 arXiv
[16]

Towards UA Vs in construction: ad- vancements, challenges, and future directions for monitoring and inspection,

Liang, Han and Lee, Seong-Cheol and Bae, Woosung and Kim, Jeongyun and Seo, Suyoung, “Towards UA Vs in construction: ad- vancements, challenges, and future directions for monitoring and inspection,” Drones, vol. 7, no. 3, pp. 202, 2023

2023
[17]

Planning and executing con- struction inspections with unmanned aerial vehicles,

Freimuth, Henk and K ¨onig, Markus “Planning and executing con- struction inspections with unmanned aerial vehicles,” Automation in Construction, vol. 96, pp. 540–553, 2018

2018
[18]

Integrating Drone Imagery and AI for Improved Construction Site Management through Building Information Modeling,

Choi, Wonjun and Na, Seunguk and Heo, Seokjae, “Integrating Drone Imagery and AI for Improved Construction Site Management through Building Information Modeling,” Buildings, vol. 14, no. 4, pp. 1106, 2024

2024
[19]

Vision meets drones: A challenge,

Zhu, Pengfei and Wen, Longyin and Bian, Xiao and Ling, Haibin and Hu, Qinghua, “Vision meets drones: A challenge,” arXiv preprint arXiv:1804.07437, 2018

2018 arXiv

Pith tools

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

[1] [1]

Mobile Manipulator for Autonomous Localization, Grasping and Precise Placement of Construction Material in a Semi-Structured Environment,

ˇStibinger, Petr and Broughton, George and Majer, Filip and Rozsyp´alek, Zden ˇek and Wang, Anthony and Jindal, Kshitij and Zhou, Alex and Thakur, Dinesh and Loianno, Giuseppe and Krajn ´ık, Tom´aˇs and Saska, Martin, “Mobile Manipulator for Autonomous Localization, Grasping and Precise Placement of Construction Material in a Semi-Structured Environment,”...

work page 2021

[2] [2]

Perceptive Model Predictive Control for Continuous Mobile Manipulation,

Pankert, Johannes and Hutter, Marco, “Perceptive Model Predictive Control for Continuous Mobile Manipulation,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 6177-6184, 2020

work page 2020

[3] [3]

Large-scale 3D printing by a team of mobile robots,

Zhang, Xu and Li, Mingyang and Lim, Jian Hui and Weng, Yiwei and Tay, Yi Wei Daniel and Pham, Hung and Pham, Quang-Cuong, “Large-scale 3D printing by a team of mobile robots,” Automation in Construction, vol. 95, pp. 98–106, 2018

work page 2018

[4] [4]

Tie a knot: human–robot cooperative workflow for assembling wooden structures using rope joints,

Mitterberger, Daniela and Atanasova, Lidia and D ¨orfler, Kathrin and Gramazio, Fabio and Kohler, Matthias, “Tie a knot: human–robot cooperative workflow for assembling wooden structures using rope joints,” Construction Robotics, vol. 6, no. 3, pp. 277–292, 2022

work page 2022

[5] [5]

Accurate and Adaptive in Situ Fabrication of an Undulated Wall Using an on-Board Visual Sensing System,

Lussi, Manuel and Sandy, Timothy and D ¨orfler, Kathrin and Hack, Norman and Gramazio, Fabio and Kohler, Matthias and Buchli, Jonas, “Accurate and Adaptive in Situ Fabrication of an Undulated Wall Using an on-Board Visual Sensing System,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), 2018, pp. 3532-3539

work page 2018

[6] [6]

Robot for automatic waste sorting on construction sites,

Chen, Xinxing and Huang, Huaiyang and Liu, Yuxuan and Li, Jiqing and Liu, Ming, “Robot for automatic waste sorting on construction sites,” Automation in Construction, vol. 141, pp. 104387, 2022

work page 2022

[7] [7]

Intelligent demolition robot: Structural statics, collision detection, and dynamic control,

Mu, Zonggao and Liu, Liyuan and Jia, Lihui and Zhang, Luyang and Ding, Ning and Wang, Chengjiang, “Intelligent demolition robot: Structural statics, collision detection, and dynamic control,” Automa- tion in Construction, vol. 142, pp. 104490, 2022

work page 2022

[8] [8]

Robots on construction sites: The potential and challenges of on-site digital fabrication,

Loveridge, Russell and Coray, Tanja, “Robots on construction sites: The potential and challenges of on-site digital fabrication,” Science Robotics, vol. 2, no. 5, pp. eaan3674, 2017

work page 2017