Year: 2026
Type: article
Abstract: A YOLOv8m detector fine-tuned only on synthetic PyBullet images (from COCO-pretrained weights) finds a requested fruit, and a simulated Franka Panda arm picks it up by aiming from the camera's depth image. The detector reached 0.995 mAP@50 and correctly classified all 149 visible targets. Grasping from depth alone succeeded in 92% of banana, 82% of orange and 50% of pear trials. All but one failure was a missed grasp, which makes grasping, not vision, the remaining limit.
Author: Powell, Roderick
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237388116
