ACT-based autonomous insertion of soil sensors using a robotic manipulator for agricultural environmental monitoring

Abstract

To address the problems of low efficiency and poor consistency in manual deployment of soil sensors for smart agricultural environmental monitoring, an ACT-based autonomous insertion method using a robotic manipulator is proposed. An experimental platform consisting of a master-slave robotic system, a dual-view vision system, and the LeRobot framework was developed. Demonstration data were collected to construct an imitation learning dataset containing wrist-view images, side-view images, and robot state information. The ACT model was employed to jointly learn visual features and robot states, enabling continuous manipulation skills including grasping, transporting, aligning, and inserting. A total of 69 episodes containing 62,031 image frames were collected, and the model was trained for 30,000 iterations. Experimental results showed that the proposed method successfully accomplished autonomous insertion of soil temperature and moisture sensors, achieving a success rate of 75.0% over 20 test trials. The results demonstrate the feasibility of applying ACT-based imitation learning to autonomous deployment of agricultural sensing nodes and provide a new solution for intelligent agricultural environmental monitoring.