PaperlayerView exact revision on arXiv

Source-linked AI summary

Design and Implementation of an Ultra-Low-Cost Wall-Climbing Robot for Infrastructure Crack Detection

Mrinmoy Modak, Supreyo Chakravorty Pretom, Shourv Tarafder, Daniel S. Drew

arXiv:2609.26130v1cs.RO

TL;DR

Crack inspection on walls, ceilings, and narrow passages is difficult to access, while manual inspection is labor-intensive and potentially hazardous. The paper combines a low-cost wall-climbing robot with a staged deep-learning workflow; its funnel-shaped geometry reduces duty cycle by about 44%, and EfficientNet-B0 outperforms the other reported models overall.

  • Problem

    Crack inspection on vertical walls, ceilings, and narrow passages is constrained by access and viewing difficulties, while conventional inspections are subjective, labor-intensive, and potentially hazardous.

    Source: §I · p. 1

  • Method

    The system pairs a fan-based wall-climbing robot and real-time video capture with YOLOv8 wall localization followed by crack analysis using YOLOv8, CNN, or EfficientNet-B0.

    Source: §I · p. 1, F. Video Processing · p. 5

  • Results

    About 44% lower duty cycle is reported for the funnel-shaped geometry, and EfficientNet-B0 outperforms YOLOv8 and CNN in overall performance.

    Source: §IX · p. 8

  • Takeaways & Limitations

    The approximately $25 robot and open collection of 40,172 labeled crack and non-crack images are presented for low-cost infrastructure inspection and further research.

    Source: §I · p. 2

Abstract

from arXiv · show

Crack detection is a crucial process to ensure the safety and longevity of buildings and other infrastructure. In this paper, we developed a low-cost, automated crack detection robot that leverages CNN, EfficientNet-B0, and YOLOv8 for efficient identification of cracks in concrete surfaces with a curated crack image dataset introduced to support training and evaluation. YOLOv8's real-time object detection enhances crack localization, while CNN and EfficientNet-B0 provide binary classification, ensuring high precision and recall. The system consists of two stages. In the first stage, YOLOv8 detects and localizes wall regions from the video frame, and the bounding boxes are cropped. The second stage performs crack detection using one of three models by analyzing the cropped regions. Cost-effective approaches are also taken for robot design. The robot features a fan-based negative pressure adhesion system, a 4-wheeled skid-steering drive, and an ESP32-CAM for real-time image capture. Its lightweight 3D-printed chassis ensures stability, allowing it to navigate both walls and ceilings while capturing images for crack analysis. Unlike conventional wall-climbing robot designs, this robot incorporates a funnel-shaped body that enhances negative pressure generation and achieves a 44% reduction in duty cycle, significantly lowering power consumption. By combining low-cost hardware with a deep learning pipeline, our system provides a scalable, efficient, and accessible solution for real-time infrastructure inspection at an approximate total cost of $25, with a lightweight web application enabling smartphone-based control. This affordability makes the system more suitable for the developing world, where infrastructure inspection is often limited by budget constraints, labor intensity, and safety risks.

I. INTRODUCTION

The paper addresses difficult, hazardous, and costly manual crack inspection with a low-cost wall-climbing robot and deep-learning workflow. Its funnel-shaped chassis is designed to improve adhesion and energy efficiency, alongside an open labeled crack dataset.

I. INTRODUCTION

  • Crack inspection in vertical walls, ceilings, and narrow passages is difficult because of access and viewing constraints, while traditional inspections are subjective, labor-intensive, and potentially hazardous.The paper also notes that many existing deep-learning systems require high-end hardware and incur substantial computational costs.Source: §I · p. 1
  • A fan-based negative-pressure robot uses four-wheel drive and an ESP32-CAM for real-time capture, with video processed by YOLOv8, EfficientNet-B0, and a CNN.The lightweight 3D-printed chassis is described as stable and agile in complex environments.Source: §I · p. 1
  • The funnel-shaped chassis passively strengthens negative-pressure adhesion and stability while reducing air leakage, increasing airflow velocity, and lowering internal static pressure.These design effects are intended to preserve energy efficiency on inclined surfaces, vertical walls, and ceilings.Source: §I · p. 1
  • The approximately $25 transportable system records real-time video and is paired with an open collection of 40,172 labeled crack and non-crack photos.The images span lighting and material conditions; lightweight deep-learning techniques are described as suitable for consumer-grade computers.Source: §I · p. 2

II. BACKGROUND AND RELATED WORKS

Prior work identifies challenges in maintaining adhesion, mobility, and reliable sensing during wall inspection, while crack-detection models face data and computation constraints. The paper situates its negative-pressure robot within a gap in affordable, scalable inspection systems.

A. Wall-Climbing Robots

  • Wall-climbing robots must balance stable adhesion with controlled mobility across surfaces that vary in texture, material, inclination, and friction.Prior approaches include ducted fans, self-compliant tracks, and bio-inspired adhesion.Source: A. Wall-Climbing Robots · p. 2
  • Vibration, optical distortion, and lighting variation can degrade inspection accuracy, motivating control approaches for stability, localization, mapping, and coordinated operation.Related work includes camera-pose alignment, behavior-based control, path-finding, drone inspection, and rock-climbing platforms.Source: A. Wall-Climbing Robots · p. 2

B. Deep Learning Approaches for Crack Detection

  • CNN methods address varied textures and lighting, but data scarcity and computationally demanding models can limit real-time or resource-constrained use.Prior approaches include multi-scale dilated convolutions, UAV-assisted CNNs, ConvNeXt V2, and attention-enhanced CNNs.Source: B. Deep Learning Approaches for Crack… · p. 2
  • Many inspection platforms rely on expensive sensors, processing units, high-power actuation, or intricate mechanics, and few combine energy-conscious design, scalable sensing, and passive adhesion efficiency.This work uses a negative-pressure approach in a low-cost, deployable architecture to address accessibility and scalability constraints.Source: B. Deep Learning Approaches for Crack… · pp. 2–3

III. SYSTEM ARCHITECTURE AND MECHANISM

The robot combines a fan-based negative-pressure adhesion system with a lightweight, four-wheel chassis and real-time camera capture. Its funnel-shaped inlet guides airflow to strengthen the pressure differential that holds the robot against a wall.

III. SYSTEM ARCHITECTURE AND MECHANISM

  • The wall-crack robot combines fan-based thrust adhesion, four-wheel drive, an ESP32-CAM for real-time image and video capture, an L298N motor driver, and a lightweight chassis for controlled movement and crack analysis.Its lightweight chassis is 3D-printed from PLA, with central clearance for airflow and camera coverage.Source: §III · p. 3
  • Ambient air enters through side inlets and passes through a convergent funnel toward a central fan, whose upward exhaust lowers pressure beneath the robot and presses it against the surface.The narrowing inlet accelerates airflow and decreases leakage; rubber tires improve traction on vertical surfaces.Source: §III · p. 3, Figure 3 · p. 3
  • The complete mechanical design files are made available online.The passage provides a link to the robot’s 3D-model files.Source: §III · p. 3

IV. DEEP LEARNING SYSTEM ARCHITECTURE

The system trains a wall-localization model before applying binary crack detection, and its dataset combines online and directly collected images. Augmentation and preprocessing produce a balanced, publicly available collection for further infrastructure-inspection research.

IV. DEEP LEARNING SYSTEM ARCHITECTURE

  • A single-class YOLOv8 model first identifies wall surfaces, after which YOLOv8, CNN, and EfficientNet-B0 models classify wall images as cracked or crack-free.The models are evaluated to identify an accurate crack-detection architecture.Source: §IV · p. 3

A. Structural Wall Crack Detection Dataset

B. Data Augmentation and Preprocessing

  • The dataset was split into training, test, and validation sets at 70:15:15, then augmented to 40,172 images: 20,050 cracked and 20,122 crack-free.Augmentation included scaling, rotation, flipping, contrast adjustment, RGB-based techniques, and cropping; it was performed after splitting to prevent data leakage.Source: B. Data Augmentation and Preprocessing · p. 3
  • The dataset is openly available to support reproducibility and further research in low-cost infrastructure inspection.The passage provides a public repository link.Source: B. Data Augmentation and Preprocessing · p. 4

C. YOLOv8 (You Only Look Once, Version 8)

YOLOv8 uses anchor-free object detection to predict wall classes and bounding boxes, and it is trained to classify wall images as cracked or non-cracked.

C. YOLOv8 (You Only Look Once, Version 8)

  • YOLOv8 predicts object classes and bounding boxes with an anchor-free detection head and updated backbone, without predefined anchors.The design is described as facilitating training, generalization, and localization accuracy, with high-speed inference for real-time use.Source: C. YOLOv8 (You Only Look Once, Version… · p. 4
  • The single-object YOLOv8 model detects wall surfaces using YOLO-format bounding-box labels and 644 × 644 pixel images.The wall-detection model is trained for 10 epochs.Source: 2) Binary Detection through YOLOv8 · p. 4

2) Binary Detection through YOLOv8:

  • For binary detection, YOLOv8 classifies resized 644 × 644 pixel wall images as cracked or non-cracked after 10 training epochs.Images are divided into training, test, and validation sets.Source: 2) Binary Detection through YOLOv8 · p. 4

D. Convolutional Neural Networks (CNN)

The CNN extracts image features through convolutional layers before classifying them, while EfficientNet-B0 uses pretrained features and fine-tuning for binary crack detection.

D. Convolutional Neural Networks (CNN)

  • The CNN processes 180×180×3 images through convolutional ReLU layers, reducing them to 45 × 45 × 64 feature maps before classification.Convolutional layers learn hierarchical visual features, including edges, textures, and shapes.Source: D. Convolutional Neural Networks (CNN) · p. 4
  • The CNN uses L2 regularization, a 128-neuron dense layer, and a sigmoid output for cracked-versus-non-cracked classification.It is trained for 25 epochs, which the authors report provided stable convergence and reliable classification performance.Source: D. Convolutional Neural Networks (CNN) · p. 4

E. EfficientNet-B0

  • EfficientNet-B0 is presented as a balance between model complexity and performance for computer vision under restricted computational resources.The paper uses it as a backbone for binary crack classification, citing its ability to achieve high accuracy with limited training data.Source: E. EfficientNet-B0 · pp. 4–5
  • EfficientNet-B0 uses ImageNet weights as a frozen feature extractor before unfreezing its last 20 layers for lower-learning-rate fine-tuning.Its inputs are resized to 224×224 pixels; two dropout layers use rates of 0.3 and 0.4.Source: E. EfficientNet-B0 · p. 5

F. Video Processing

The video pipeline analyzes camera frames in localized patches, then labels cracked regions; YOLOv8’s reported confusion matrix shows few errors, while its training curves indicate validation-loss fluctuations.

F. Video Processing

  • Each frame is split into four non-overlapping patches, and a patch below the 0.5 predicted-probability threshold is labeled cracked with its coordinates recorded.The pipeline first checks for a wall with single-object YOLOv8 detection; wall-present frames proceed through the crack-detection models, and detected cracks are shown with a labeled bounding box.Source: F. Video Processing · p. 5

A. Crack Detection By YOLOv8 Results

B. Crack detection by CNN Results

The CNN confusion matrix reports few misclassifications, while its validation loss fluctuates despite smoothly declining training loss and high accuracy.

B. Crack detection by CNN Results

  • 2984 true positives, 2990 true negatives, 29 false negatives, and 18 false positives are reported for the CNN confusion matrix.Fig. 6(b) identifies the CNN confusion matrix.Source: B. Crack detection by CNN Results · p. 5
  • The CNN’s training loss declines smoothly, but validation loss fluctuates, suggesting slight overfitting despite high accuracy.The text attributes the validation-loss instability to lower data quality and variation, and says L2 regularization compensates for some overfitting.Source: B. Crack detection by CNN Results · p. 5
  • Table I(b) reports recall, precision, and F1-score for the CNN model.Source: B. Crack detection by CNN Results · p. 6

C. Crack detection by EfficientNet-B0 Results

The section reports crack-detection outcomes, training behavior, and real-time classification and localization. The confusion matrix records correct crack and non-crack predictions with relatively few errors.

C. Crack detection by EfficientNet-B0 Results

  • 2,406 true positives and 2,350 true negatives were reported, alongside 96 false negatives and 23 false positives.These are the confusion-matrix counts for crack detection.Source: C. Crack detection by EfficientNet-B0… · p. 6
  • Training and validation losses declined smoothly, accuracy remained high, and the model showed less overfitting despite limited dataset variation.The passage points to Fig. 7(III) for the accuracy and loss curves and Table I(c) for EfficientNet-B0 recall, precision, and F1-score, without supplying those metric values.Source: C. Crack detection by EfficientNet-B0… · p. 6
  • The real-time system classifies crack and non-crack surfaces and localizes crack regions with green bounding boxes.Source: Figure 8 · p. 6

D. Real-time Crack Analysis

The system processes incoming video for real-time crack detection and reports robot mobility across surface orientations. Its funnel-shaped shroud reduced duty cycle on inverted surfaces, while the cost analysis describes the design as cost-efficient for real-life use.

D. Real-time Crack Analysis

  • The ESP32-CAM sends real-time video to a computer, which processes each frame and displays “Crack” or “No Crack” with bounding boxes when cracks are detected.Source: D. Real-time Crack Analysis · p. 7

VI. MOBILITY ASSESSMENT ON INCLINED AND INVERTED SURFACES

  • About 44%: on inverted surfaces, the funnel-shaped shroud reduced duty cycle from 90% to about 50% compared with the same chassis without the funnel.The passage associates this reduction with a significant decrease in power consumption.Source: §VI · p. 7

VII. COST ANALYSIS

  • The cost analysis concludes that the robot is highly cost-efficient and suitable for real-life applications in the developing world.The cited text says Table II gives the cost breakdown and Table III compares the suggested robot with prior robots, but does not state their values.Source: §VII · p. 7, Table III · p. 7
  • The design attributes its cost-effectiveness to locally available products and a funnel-shaped negative-pressure mechanism that reduces the need for high-power vacuum pumps or complex seals.Source: §VII · p. 7

VIII. FUTURE WORK

The conclusion reports lower duty cycle with the funnel-shaped geometry and stronger overall performance from EfficientNet-B0 than from YOLOv8 and traditional CNN models. Future work proposes testing additional architectures and adding autonomous path planning.

VIII. FUTURE WORK

  • Future work proposes investigating ResNet50 for feature extraction and U-Net for fracture-region segmentation, as additions to or replacements for current models.Source: §VIII · p. 7
  • The authors propose integrating RRT* or other path-planning algorithms so robots could detect cracks during autonomous traversal.They describe this integration as a route toward fully automated inspection and extensive infrastructure monitoring.Source: §VIII · p. 7

IX. CONCLUSION

  • EfficientNet-B0 outperformed YOLOv8 and traditional CNN models in overall performance, while the funnel-shaped geometry reduced duty cycle by about 44%.The conclusion also says image patching enables more accurate feature learning and localized analysis, and that limiting detection to pertinent regions lowers computational complexity.Source: §IX · p. 8
Loading 2609.26130v1…