The identification of leaf water content in avocados can help to improve the irrigation regimes, thus alleviating the great water consumption of the avocado production. The classification of leaves according to their dehydration stage can be used as an early assessment tool for leaf water content identification. This study focuses on classifying avocado leaves based on their dehydration stage, using two approaches: hyperspectral reflectance values with classic machine learning models, and multi-spectral images with deep convolutional neural networks. The classic machine learning algorithms trained include random forest, k-nearest neighbours, support vector machine, bagging, decision tree, reduced error pruning tree, simple classification and regression tree, Bayesian networks, logistic model tree, linear discriminant analysis, multinomial logistic regression, and multilayer perceptron. The deep learning approach employed convolutinal neural network, specifically ConvNeXt-Tiny, MobileNetV3-Large, and ResNet18. The models were evaluated using common metrics such as accuracy, precision, f1 score, receiving operating characteristic curves, area under curve and confusion matrices. Among the machine learning models, linear discriminant analysis achieved the highest accuracy of 0.98. Meanwhile, ResNet18 outperformed the other CNNs, achieving an accuracy of 0.99 on the validation dataset. These results highlight the effectiveness of the proposed methods for classifying avocado leaves based on their dehydration level.

Controlling the locomotion of autonomous legged robots requires sophisticated techniques to handle system constraints, often resulting in high computational demands that limit real-time implementation. This article presents a computationally efficient path tracking algorithm for legged robots, leveraging their kinematic model and built-in low-level motor control. Unlike traditional model predictive control (MPC), the proposed method simplifies the framework by using a single-step prediction window, achieving performance comparable to MPC with longer horizons, while significantly reducing computational overhead. Comparative analysis against a Nonlinear MPC with and without integral mode, a linear time-varying LQR, and a reinforcement learning agent demonstrates that the proposed algorithm achieves similar or superior tracking performance with much lower computational requirements and tuning. This makes the approach well-suited for real-time applications in resource-constrained robotic systems. In addition, the method facilitates fast prototyping for industrial applications, as performance and stability can be established using simple kinematic and path parameters. Our controller simplifies design requiring almost no tuning obtaining a similar performance in comparison with an NMPC with a control horizon of 20 sampling instant but with much less computational effort.

Collision avoidance is crucial for autonomous navigation systems. Many studies have addressed obstacle avoidance for single unicycles and car-like vehicles in on-road conditions. In this work, we extend the scope to generalised N-trailer vehicles, comprising a single active segment pulling multiple trailers. Unlike approaches that treat obstacles as hard constraints, we model them as soft constraints using Gaussian functions. This method maintains the convexity of the search space, reducing computational demands. However, the regions occupied by obstacles remain feasible. Thus, the Gaussian function’s amplitudes need to be carefully chosen to discourage navigation through these areas. Moreover, closed-loop stability is guaranteed by generating auxiliary references when the nominal path is occluded. The efficacy of this approach is demonstrated through simulated and field experiments with a tractor pulling two trailers. These experiments show the method’s capability to navigate around obstacles efficiently while maintaining computational efficiency, validating its practical applicability. Videos of the experiments and the implemented algorithms are available athttps://usmcl-my.sharepoint.com/:f:/g/personal/nestor_deniz_usm_cl/EtU54g1NeslNhD8V7dAeu20B0umnQa4FKiMlzThkTAXYvg?e=swEXwg. Despite the success in real-time implementation, more research is needed to address the open questions discussed at the end of this article. Note to Practitioners—This work focuses on implementing obstacle avoidance for a kind of vehicles widely used in agriculture, mining, luggage transportation, and industry. A LiDAR Velodyne VLP16, configured with its lowest rotation speed for denser point clouds, is used to scan the environment. Proper attachment of the LiDAR to the tractor’s body minimises vibration and azimuth movements, ensuring accurate obstacle detection. Obstacles are modelled as Gaussian functions to maintain the convexity and optimise computational efficiency. The Gaussian function’s amplitude should be set high enough to effectively avoid collision when density of obstacle is high. The framework uses a control horizon$N_{c}$and a prediction horizon$N_{p}$beyond the control to anticipate obstacle’s position. However, a large prediction horizons$N_{p}$is not advised when the model of the dynamic of the obstacles is not accurate.

The phenotyping of plants is becoming more relevant to effectively managing the expectations associated with a product with certified quality, enhancing profitability, and increasing field and crop productivity. Although several solutions have been proposed to characterize plants in physical and biochemical aspects, the main contributions have been related to costly dedicated instruments. This work presents an instantaneous estimator for plant functional traits by harnessing the harvested power from an electric field energy harvester (EFEH). Specifically, we establish the detected correlation between twenty vegetation indices associated with water content and the opencircuit voltage (VOC) and short-circuit current (ISC) of an EFEH assembled with natural leaves. To this end, several 10×3 cm2EFEHs were assembled using natural leaves sourced from two distinct species: Magnolia Obovata and Ravenala Madagascariensis. Each EFEH underwent a four-stage dehydration process. The primary outcome of this work is the exploration of VOCand ISCto retrieve fuel moisture content (FMC) and equivalent water thickness (EWT) based on machine learning models. The results indicated that the electrical parameter with the highest coefficient of determination was ISC, which presented an R2of up to 0.7691 and 0.7639 to retrieve FMC and EWT, respectively.

Mobile robots are widely used in agriculture, either as the form of intelligent vehicles or as monitoring platforms, where sensors are placed for the gathering of data. Overall, wheeled robots have play a crucial role in the development of new monitoring techniques, mainly due to their capabilities (their movements can resemble the ones of a tractor) or because of the wide variety found on the market. Nevertheless, another kind of mobile robots, known as legged robots, have not been explored yet in the agricultural context, specifically as platforms for data gathering and eventually, for interaction with the growing. In this work, we compare the capabilities of a legged robot against a wheeled one following previously published guidelines for performance evaluation of both kind of robots. In particular, we evaluate the odometry responses and the energy consumption of the robots when performing previously known manoeuvres. Moreover, considering that in the agricultural context is feasible to find different kind of terrains, we performed our evaluation under grass, gravel and pavement, both even and uneven terrains, for a circle-shaped, closed S-shaped and square-shaped, of different sizes. Overall, the legged robot showed a more uniform response in the power consumption behaviour despite the kind of terrain, but the wheeled robot showed better results from a localization perspective. The analysis provided herein serve as the basis for future usage of legged robots in the agricultural context.

Bare-earth extraction in forested regions has been considered challenging because of the lack of ground point information. In these regions, vision systems cannot capture any information about ground points under the canopy. Thus, the challenge of generating a digital terrain model by cameras increases. Nevertheless, one might alleviate ground filtering using vegetation’s features (e.g., chlorophyll). In this regard, this article evaluated two machine-learning approaches [i.e., conditional random field (CRF), artificial neural network (ANN)] for generating digital terrain models when biophysical or biochemical features of vegetation are given. Terrain models were generated from multispectral image-based point clouds. A fivefold cross-validation methodology evaluated the CRF and ANN. The point clouds were retrieved from two study areas at different illumination and flight altitudes. Vegetation features were computed as vegetation indices from the multispectral point clouds. Results suggested that by using these indices, the classification of ground points could be enhanced. In particular, the vegetation indices that yielded the best outcomes were normalized difference vegetation index, green NDVI, and modified chlorophyll absorption reflectance index. Moreover, it was shown that CRF generates elevation models more smoothly than a triangular irregular network method. Thus, a CRF could be promising for classifying ground points in forested regions using geometric and vegetation features from a photogrammetric point cloud.

Forest regions are still considered as complex environments for measuring terrain information by aerial surveys. Tree canopies occluded most of the ground surface and limit sensors’ capabilities for capturing ground data. For instance, vision systems (e.g., cameras) cannot record any information about the ground below the canopy. This lack of knowledge might decrease the accuracy of aerial surveys’ products such as digital terrain models (DTMs). Therefore, to outperform the ground surface knowledge, we proposed a method for computing the elevation of occluded ground points using canopy data and estimated tree height. To this aim, individual tree crowns are identified from a 3-D point cloud, retrieved by discrete light detection and ranging (LiDAR) sensors (i.e., Riegl LMSQ560, ALTM 3100, and Riegl LMSQ5600). Then, ground elevation is predicted by taking advantage of the logarithmic relationship between crown diameter (CD) and tree height. Results have shown that the mean, minimum, and maximum ground elevation errors are about 3.01, 1.28, and 6.38, respectively. Conversely, if one attempts to determine a ground surface without the proposed method, the mean elevation difference is about 11.68.

Additive manufacturing is a disruptive technology that enables the efficient construction of lighter and stronger concrete structures. In general, fabrication procedures are produced by industrial 3D printers, which constantly deposit concrete to ensure high building standards and reduce potential deformations in generated components. However, the deposition rate of building materials remains an empirical and heuristic procedure that depends on prior knowledge of the model. This work introduces a methodology to automatically detect deformations in printed layers by analyzing the 3D characterization of concrete structures. To this end, the performance of a monocular camera, LiDAR, and LiDAR-camera is studied according to point cloud density and 3D map reconstruction. In addition, a portable ground-based system for detecting possible deformations is conceived, manufactured, and experimentally tested. Empirical findings show that the proposed system is capable of detecting printed layer variation with a low error of 0.3%, revealing that the low-cost sensors can be an autonomous and highly reliable solution for deformation detection in concrete structures.

In general, proposed solutions for LiDAR-based localization used in autonomous cars require expensive sensors and computationally expensive mapping processes. Moreover, the global localization for autonomous driving is converging to the use of maps. Straightforward strategies to reduce the costs are to produce simpler sensors and use maps already available on the Internet. Here, an analysis is presented to show how simple can a LiDAR sensor be without degrading the localization accuracy that uses road and satellite maps together to globally pose the car. Three characteristics of the sensor are evaluated: the number of range readings, the amount of noise in the LiDAR readings, and the frame rate, with the aim of finding the minimum number of LiDAR lines, the maximum acceptable noise and the sensor frame rate needed to obtain an accurate position estimation. The analysis is performed using an autonomous car in complex field scenarios equipped with a 3D LiDAR Velodyne HDL-32E. Several experiments were conducted reducing the number of frames, the number of scans per 3D point-cloud and artificially adding up to 15% of error in the ray length. Among other results, we found that using only 4 vertical lines per scan and with an artificial error added up to 15% of the ray length, the car was capable to localize itself within 2.11 meters error average. All experimental results and the followed methodology are explained in detail herein.

Human-robot interaction (HRI) techniques has become important in the resolution of problems that can not be automated completely. Moreover, the design of strategies based on such interactions can help to compensate the agents respective limitations. This work aims the development of a HRI system, which is based on driver sleepiness to improve safety in a Renault Twizy vehicle. For this purpose, eye behavior measurements are extracted using a properly machine learning algorithm that can operate under varying illumination levels. Perclos and blink frequency measures are considered to estimate diver sleepiness, and based in such parameters, the vehicle speed is limited, which may help decrease accidents rate. The presented architecture can improve the security conditions of navigation for both users and pedestrians. The results show that it is possible to implement HRI strategies based on cognitive factors in urban areas to prevent fatal accidents.

Extract and recognize face features can become a difficult problem, especially in environments with dynamic illumination conditions. For example, changing faces position respect to the camera and varying intensity of the light source, among others. Trying to mitigate illumination variation effects have been studied using different approaches, but a comparison between them and their characteristics such as processing times using a classifier is still needed. This is important to try to find a properly algorithm that can fulfill the demanding requirements for some cognitive applications. In this work, an illumination invariant face feature recognition using dual-tree complex wavelet transform is presented. A validation and testing of the proposed approach is performed using Yale B faces dataset, showing that we can obtain 90.7% to 98.5% recognition rates on the dataset depending of the illumination level with the proposed method. Additionally, a comparison between 21 other illumination normalization methods using the same classification approach is presented. Finally, an online algorithm is implemented and tested on real environments under varying lighting conditions, which is capable to recognize subject faces, and their eyes and mouth status. In particular, the on-line results of the proposed approach show recognition rates for eye blink detection from 84.1% to 90.4% on 55[ms], which may be useful for time demanding applications such as sleepiness detection.

Reducing material spillage by robotic mining mobile manipulators, such as front-end loaders, is necessary to improve mining operations. To this end, the present work proposes an approach to reduce disturbances on the end-effector induced by the terrain and propagated through the wheels and arm links of the machine. The proposed approach is based on an H∞control strategy that includes a feedforward action, computed using the pitch rate of the mobile base, and considers the hydraulic arm dynamics, as well as the reaction forces in the contact points of the mobile base, which is modeled as a floating body with non-permanent ground contacts. Alternative control schemes based on the classic proportional-derivative (PD) control, and the Active Disturbance Rejection Control (ADRC), with and without feedforward action, were also implemented and experimentally evaluated using a semiautonomous Cat®262C compact skid-steer loader equipped with inclination and inertial sensors. The proposed method reduces disturbances by at least 70% when climbing ramps at 25% of the machine's maximum speed, and by at least 20% when driving over speed bumps which produce disturbances similar to that caused by stones. The proposed disturbance attenuation strategy should help reducing the spillage of material when driving over mounds, inclines or spilled rocks, especially considering that even if existing autonomous machines are able to drive with little operator supervision along mining galleries, they are often unable to avoid disturbing material on the ground or the characteristic unevenness of mining terrains.

Nowadays, agricultural and mining industry applications require saving energy in mobile robotic tasks. This critical issue encouraged us to enhance the performance of path tracking controllers during manoeuvring over slippery and rough terrains. In this scenario, we propose probabilistic approaches under machine learning schemes in order to optimally self-tune the controller. The approaches are real time implemented and tested in a mining machinery skid steer loader Cat® 262C under gravel and muddy terrains (and their transitions). Finally, experimental results presented in this work show that the performance of the controller enhances up to 20% (average) without compromising saturations in the actuators.

Performance of path tracking controllers for mobile robots is usually constrained to the tuning of a set of gains within previously known boundaries. In industrial applications, controller tuning is a time consuming task, which is especially critical when the gain values have a considerable impact on the use of the robot's energy resources. In this paper, we propose a set of approaches to automatically improve the performance of well-known path tracking controllers used for agricultural applications of automated agricultural machinery. The effect of the proposed strategy can be better appreciated during the transient response of the robot when it has to maneuver along a non-kinematically compatible path. The results presented in this work show that the robot's path tracking performance is improved up to 15% using the proposed approaches.

Vehicle localization in large-scale urban environments has been commonly addressed as a map-matching problem in the literature. Generally, the maps are 2D images of the world where each pixel covers a part of it. However, building maps for large-scale urban environments requires driving the vehicle along the desired path at least once. In order to simplify this task, in this work, we propose a new localization system that uses satellite aerial map-images available on the Internet to localize a vehicle in a complex urban environment. Satellite aerial map-images are compared against re-emission maps built from the infrared reflectance information of the vehicle's LiDAR. Normalized Mutual Information (NMI) is used to compare re-emission and aerial map images. A Particle Filter Localization strategy is applied for vehicle's localization. As a result, the system has an accuracy of 0.89m in a test course with 6.5km. Our system can be used continuously without losing track, and it works even in dark and partially occluded areas.

The present work models the dynamics of general skid-steer mobile manipulators using the formalism and tools of the spatial vectors algebra introduced by Featherstone. The model built is validated using inertial measurements obtained during field tests with a compact skid-steer loader. The paper demonstrates the benefits of using the spatial vector algebra formulation, showing that this modeling approach allows to integrate traction forces and study the arm-vehicle, as well as vehicle-ground interactions in a single model. This feature is not possible with many other of the existing modeling approaches and simulation tools, thus opens the way to research on mechanically more complex robot designs and their controllers. It is to be noted that most of the existing models and simulations of mobile manipulators consider two-wheeled differentially driven bases and avoid accurate models of skid-steering bases because of the complexity of simulating wheels that skid while rolling. However, skid-steer traction is common in most of the industrial construction and mining machinery because of their simpler mechanics, high reliability, and better mobility in rough terrains. Hence, the development of physically accurate models of skid-steer manipulators is fundamental. We chose to validate the model using a Cat®262C compact-skid steer loader instead of a small mobile manipulator common in robotics research laboratory to highlight the usefulness of the presented model and the spatial vector algebra approach.

This paper proposes a solution to a door crossing problem in unknown environments for an autonomous wheelchair. The problem is solved by a dynamic path planning algorithm implementation based on successive frontier points determination. An adaptive trajectory tracking control based on the dynamic model is implemented on the vehicle to direct the wheelchair motion along the path in a smooth movement. An EKF feature-based SLAM is also implemented on the vehicle which gives an estimate of the wheelchair pose inside the environment. The SLAM allows the map reconstruction of the environment for future safe navigation purposes. The entire system is evaluated in a real time simulator of a robotic wheelchair.