Autonomous navigation system in various greenhouse scenarios based on improved FAST-LIO2

Abstract

The development of phenotypic detection robots suitable for semi-structured greenhouses is of significant importance for accelerating crop breeding, particularly in the screening of advantageous germplasm resources. However, the diversity of greenhouse structures and the limitations of GPS signals pose challenges to the autonomous navigation of robots. In this study, a system with autonomous navigation, voice interaction, and adaptive data acquisition was developed for strawberry germplasm resources. To reduce the drift of the global map on the z-axis and improve consistency, ground constraints and stable triangle descriptor (STD) loop closure detection were incorporated into the fast direct light detection and ranging inertial odometry (FAST-LIO2) framework. In addition, the improved FAST-LIO2 and Kalman filter were utilized to provide poses, achieving precise and continuous localization of the robot. To improve flexibility, the demonstrated path was utilized as the global path. Besides, the system integrated adaptive data acquisition and voice control modules, enabling the automatic collection of target plant data and variety information while enhancing human–computer interaction performance. The system achieved high-precision navigation across different scenarios, speeds, and motion states. Even in the state of lowest accuracy during row change, the standard deviation (SD) of the total deviation remained below 2.6 cm, the root mean square error (RMSE) was less than 5.9 cm, and the average deviation (AD) was below 5.3 cm. In terms of heading deviation, the SD was below 1.8°, the RMSE was less than 3.8°, and the AD was below 3.4°. Moreover, the success rate of target plant detection reached over 98 %. This system facilitated the construction of phenotypic analysis models, assisting breeders in variety management and demonstrating application potential in greenhouse phenotypic detection.