A robust agricultural LiDAR-inertial SLAM system with heterogeneous registration and neural noise adaptation

To address the challenges of unstructured terrain, repetitive geometric structures, and significant inertial disturbances in agricultural environments, this paper proposes a robust LiDAR-inertial simultaneous localization and mapping (SLAM) system for agricultural robots. The proposed system is built upon the LIO-SAM framework and incorporates two key improvements. Firstly, a vegetation–structure heterogeneous registration strategy is introduced. By leveraging point cloud intensity information, rigid structures and flexible vegetation are separated, and two parallel registration pipelines are constructed: a structure-oriented channel based on normal-distribution modeling and a vegetation-oriented channel based on geometric features. Pose estimation is then achieved through confidence-weighted fusion, improving registration robustness in scenes with repetitive structures. Secondly, a neural noise adapter (NNA) is developed to dynamically model the time-varying characteristics of IMU noise using a long short-term memory (LSTM) network. By fusing raw IMU data, instantaneous noise statistics, and terrain undulation features, the adapter predicts the pre-integration covariance scaling factor in real time, effectively suppressing pose drift caused by rugged terrain. Comparative experiments conducted on public datasets, a Gazebo agricultural simulation environment, and real orchard scenarios demonstrate the proposed method’s advantages in localization accuracy, and generalizability, providing reliable perception support for precision agricultural robotics.