publications

2025

1. 

   Comparative Analysis of LiDAR Inertial Odometry Algorithms in Blueberry Crops

   Ricardo Huaman, Clayder Gonzalez, and Sixto Prado

   In , 2025

   Abs DOI

   In recent years, LiDAR Odometry (LO) and LiDAR Inertial Odometry (LIO) algorithms for robot localization have considerably improved, with significant advancements demonstrated in various benchmarks. However, their performance in agricultural environments remains underexplored. This study addresses this gap by evaluating five state-of-the-art LO and LIO algorithms—LeGO-LOAM, DLO, DLIO, FAST-LIO2, and Point-LIO—in a blueberry farm setting. Using an Ouster OS1-32 LiDAR mounted on a four-wheeled mobile robot, the algorithms were evaluated using the translational error metric across four distinct sequences. DLIO showed the highest accuracy across all sequences, with a minimal error of 0.126 m over a 230 m path, while FAST-LIO2 achieved its lowest translational error of 0.606 m on a U-shaped path. LeGO-LOAM, however, struggled due to the environment’s lack of linear and planar features. The results underscore the effectiveness and potential limitations of these algorithms in agricultural environments, offering insights into future improvements and adaptations.

2024

1. 

   Performance Evaluation of the ROS Navigation Stack Using LeGO-LOAM

   Ricardo Huaman, Clayder Gonzalez, and Sixto Prado

   In Proceedings of the 9th Brazilian Technology Symposium (BTSym’23), Cham, 2024

   Series Title: Smart Innovation, Systems and Technologies

   Abs DOI

   In recent years, autonomous mobile robots have been increasingly employed in various domains due to their ability to navigate unpredictable environments with limited human intervention. The Robot Operating System (ROS) Navigation Stack is a contributor to this development. The present research assessed the effectiveness of the Lightweight and Ground-Optimized Lidar Odometry and Mapping (LeGO-LOAM) when integrated into the ROS Navigation Stack for a skid-steering mobile robot operating in two distinct settings. Performance was benchmarked against well-established algorithms: Gmapping handling mapping and AMCL addressing localization. The mapping experiment demonstrated that both Gmapping and LeGO-LOAM generated precise occupancy grid maps, with LeGO-LOAM providing finer environmental details due to its 3D point cloud projection. In navigation tasks, the Absolute Pose Error (APE) metric revealed AMCL’s marginally superior performance in the first environment, achieving an RMSE value of 1.637 against LeGO-LOAM’s 1.696. Conversely, in the second environment, LeGO-LOAM outperformed AMCL by recording a lower RMSE of 1.976, compared to AMCL’s 3.163. These findings support LeGO-LOAM’s viability as an alternative tool suited to localization and mapping tasks within the ROS Navigation Stack and emphasize the significance of selecting algorithms tailored to specific environments to achieve optimal performance.

2023

1. 

   Linear Quadratic Regulator (LQR) Control for the Active Suspension System of a Four-Wheeled Agricultural Robot

   Jheremy Andre Bazan Quispe, Ricardo Jesus Huaman Kemper, and Sixto Ricardo Prado Gardini

   In 2023 IEEE XXX International Conference on Electronics, Electrical Engineering and Computing (INTERCON), Lima, Peru, 2023

   Abs DOI

   As the adoption of robotics within the agricultural industry increases, the issue of maintaining stability and precise task execution on irregular terrain has emerged as a critical challenge. To address this, this study introduces a control strategy for enhancing the suspension system of an agricultural mobile robot. A dynamic model, characterized in the state-space representation, was formulated to encapsulate the robot’s suspension dynamics. This model was then interfaced with a Linear Quadratic Regulator (LQR) for optimized control. Performance evaluations were conducted using MATLAB/Simulink, and two contrasting road profiles were employed to simulate agricultural terrains. When compared to the passive counterpart, the active LQR-controlled suspension presented substantial improvements. Specifically, for the initial road profile, the roll angle, pitch angle, and vertical acceleration showed improvements of 40.68%, 39.47%, and 37.73%, respectively. On the second road profile characterized by depressions, these metrics improved by 52.02%, 66.97%, and 71.03%. Consequently, the proposed LQR control strategy enhances the robot’s stability and precision on challenging agricultural terrains by reducing oscillations and minimizing stabilization time.

2. 

   Deep Learning-based Segmentation and Classification System for Artichoke Seedling Grading

   Percy Brayam Cubas Muñoz, Ricardo Jesus Huaman Kemper, Erick Martin Fiestas Sorogastua, and 1 more author

   In 2023 IEEE XXX International Conference on Electronics, Electrical Engineering and Computing (INTERCON), Lima, Peru, 2023

   Abs DOI

   This study introduces a strategy for the automated grading of artichoke seedlings, a critical task in the transplantation process. The proposed strategy leverages two distinct perspectives, vertical and horizontal, to capture images of the seedlings using depth cameras. These images are then processed through segmentation modules to extract specific morphological attributes such as width, height, and leaf area. A feature vector composed of these attributes is evaluated by a Multi-Layer Perceptron (MLP) to classify the seedlings into two categories: ‘good’ and ‘not good’. Two deep learning models, Mask RCNN and YOLOv8, were employed for detection and instance segmentation tasks. Various backbones were experimented with, and the most effective ones were identified for both models. An MLP was developed for the classification stage, achieving an accuracy of 85.19% with specific training parameters. This research contributes a foundational framework for automated seedling grading, with potential applications and improvements in the field of agricultural automation.

3. 

   Artificial vision strategy for Ripeness assessment of Blueberries on Images taken during Pre-harvest stage in Agroindustrial Environments using Deep Learning Techniques

   Percy Brayam Cubas Muñoz, Ricardo Jesus Huaman Kemper, and Sixto Ricardo Prado Gardini

   In 2023 IEEE XXX International Conference on Electronics, Electrical Engineering and Computing (INTERCON), Lima, Peru, 2023

   Abs DOI

   This article introduces a methodology aimed at detecting and categorizing blueberries within an agro-industrial farm. The methodology employs a combination of YOLOv8 and a custom convolutional neural network. The process initiates with the creation of a dataset used for training and validating the proposed approach. Subsequently, the manual generation of masks for blueberries in the captured images is carried out, ensuring precise labeling that corresponds to five distinct ripeness levels. A model is developed to detect and classify the blueberries based on their respective ripeness levels. The evaluation of these models demonstrates a notable accuracy of 91.14% in the task of classifying the blueberries.

2022

1. 

   Autonomous Navigation of a Four-Wheeled Robot in a Simulated Blueberry Farm Environment

   Ricardo Jesus Huaman Kemper, Clayder Gonzalez, and Sixto Ricardo Prado Gardini

   In 2022 IEEE ANDESCON, Barranquilla, Colombia, 2022

   Abs DOI

   Fruit sampling for blueberry harvesting projections are labor-intensive and if not properly done cause delay and economic loss throughout the entire supply chain. The use of robots could be a promising tool to address these issues. However, field tests for robots in agricultural environments are not always readily available. This paper presents a simulation for a ROS based four-wheeled agricultural robot which navigates autonomously through blueberry bushes rows. The autonomous navigation system comprises a Simultaneous Localization and Mapping algorithm (LeGO-LOAM), a path planning algorithm (A*) and a path tracking algorithm (an ad hoc algorithm). Its performance is assessed using the ATE metric in a simulated blueberry farm environment in Gazebo simulator. The proposed navigation system allowed the robot to navigate successfully in this environment with the highest error reaching 500 mm and 250 mm for the x and y Euclidean components respectively. The presented simulation methodology will expedite development of agricultural robots and ultimately aid in overcoming the aforementioned issues.