, 남세현, 박종천, 서현욱, 최태용
, Se Hyun Nam, Jongcheon Park, Hyunuk Seo, Taeyong Choi
Vertical farms provide structured, controlled environments for stable crop production and suit robotic automation. However, autonomous harvesting there remains challenging because visually similar fruits are densely arranged and can occlude the target fruit. We propose a Vision-Language-Action (VLA) pipeline augmented with explicit target-object information for dualarm robotic strawberry harvesting. A perception module selects the target strawberry from wrist-camera images; its 3D position is concatenated with the robot state, and its segmentation mask is overlaid on the RGB image. The pipeline, built on GR00T N1.7, was evaluated in 220 real-robot trials under four input configurations: baseline, mask only, position only, and both. In single-object trials (45 each), success rates were 75.6%, 91.1%, 88.9%, and 95.6%, respectively; only the last differed significantly from the baseline, and the augmented configurations did not differ significantly from one another. The improvement arose mainly after failed grasps: the baseline tended to regrasp at the same location, whereas target-informed policies re-approached the target. In a three-strawberry cluster with one fruit occluded in the head-camera view, policies given the 3D position succeeded in all 5 trials, versus 2 for the baseline. These results suggest that explicit target-object information improves VLA-based harvesting under occlusion within the tested layouts.
