Image-guidance for robotic harvesting of micropropagated plants

Abstract This paper describes a computer vision algorithm used to guide a laboratory robot in grasping the stems of micropropagated plants. The plants were chrysanthemums, growing in containers of agar, and were typical of plants ready for harvesting and dissection. The plants were imaged in silhouette. Pieces of stem were distinguished from other foliage by their thickness, length, angle-to-vertical, and connection downwards to the bottom of the container. These were combined to give, for each piece of stem, a measurement of the distance from the camera, and the estimated accuracy of the measurement. Experiments were conducted in which the stem segments were arranged in order of preference, depending on gripper clearance and the accuracy of the stereo measurement, and the tips of the grippers were guided to the most preferred segments. The amount of foliage affected the algorithm performance, and this varied between the plant clones. With the small-leaved clones, stems were identified with about 90% certainty, but with the large-leaved clones, very little detail could be seen in silhouette, and the performance was poor. The stereo measurement located the most preferred stems to within at least ±2 mm, and rarely resulted in the stem being missed or two stems being grasped. Failure to grasp a stem was mainly due to other foliage obstructing the grippers. The plants could not be pulled out of the container once grasped because the roots were too large. Recommendations for future work included more careful clearance checking and less bulky grippers to reduce the problems due to obstructing foliage, and improved image understanding to locate the bases of the plants, where they could be cut rather than pulled.