Navigating with Finesse: Leveraging Neural Network-based Lidar Perception and iLQR Control for Intelligent Agriculture Robotics

Abstract

Autonomous navigation has revolutionized agriculture by enabling robots to interact with the environment autonomously. This article presents an integrated system that addresses two key aspects: generalizing environments using LiDAR perception and neural networks, and achieving optimal solutions for nonlinear systems with low computational cost. To achieve precise and efficient navigation in complex agronomic environments, the system combines LiDAR sensors and deep learning methods, specifically utilizing a ResNet-based neural network architecture. LiDAR sensors provide accurate and detailed information on terrain, crops, and obstacles, while the ResNet architecture enhances perception capabilities by extracting and analyzing features from LiDAR point cloud data. For smooth and accurate trajectory following, the system employs the iLQR algorithm, which calculates control commands using an optimization-based control method for nonlinear systems. This algorithm ensures robust guidance of the robot along the desired trajectory. By integrating LiDAR perception with the ResNet-based deep learning approach and iLQR control, the system enhances the navigation capabilities of agricultural robots, resulting in reduced operational costs, increased efficiency, and minimized environmental impact.