Rapid path definition and path-tracking method of CleaningBot in stacked cage farming houses based on learning from demonstration

Abstract

Autonomous navigation within confined stacked cage farming houses (SCFHs) poses several challenges: repetitive structures negatively impact the performance of scan-matching-based simultaneous localization and mapping (SLAM); vision-based systems are impeded by weak textures from mesh cages, motion blur, and lighting variations; and the long-term reliability of fiducial markers is compromised by harsh dusty conditions. To address these challenges, this study introduces a comprehensive map-free navigation framework implemented in CleaningBot, a two-wheeled differential-drive autonomous mobile robot. Central to this framework is a novel learning-oriented algorithm for rapid path definition. By employing the learning from demonstration (LfD) method, CleaningBot captures key action points and directional angles, enabling path definition within minutes based on a single human demonstration, which is a substantial improvement over manual waypointing, a process that can take more than 50 min. For precise path execution, a feedback linearization straight-line tracking controller was used to minimize the lateral and angular deviations. This was integrated with a reactive obstacle avoidance controller to form an adaptive path and obstacle navigation controller (APONC). The system’s robustness relies on a high-precision inertial localization module, enhanced by an improved University of Michigan benchmark (UMBmark)-based odometry calibration method. Extensive experiments confirmed the performance of the system, highlighting three key achievements: (1) The robot achieved a positioning error of only 0.053 % over a 500 m trajectory, significantly outperforming common SLAM-based methods in degraded settings. (2) In real SCFH environments, trajectory reproduction achieved overlap rates of more than 99 %. (3) The APONC enabled the robot to avoid a moving person and realign with its path in 3.2 s, with a final residual lateral error of only 0.031 m. This study presents a practical, efficient, and robust navigation solution for confined agricultural environments.