그리퍼 유연 접촉층의 두께에 따른 VLA 기반 로봇 조작 성능 평가

초록

This study evaluates how the thickness of a flexible gripper contact layer affects a fine-tuned Vision-Language-Action (VLA) policy. Octo-Small is selected under a resource-constrained GPU environment and fine-tuned using 50 demonstrations. Pick-and-place experiments are conducted 20 times for each thickness of 0, 1, 2, and 3 mm. The success rates are 75.0%, 70.0%, 80.0%, and 40.0%, respectively, with more grasp attempts observed at 3 mm. These results indicate that excessive contact-layer thickness may degrade VLA-based manipulation performance by altering the gripper’s appearance and contact conditions.

키워드