Reproduce “SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation” on OpenResearch

About this paper

Kaivalya Agrawal, Md Ashiqur Rahman, Raymond A. Yeh, Zachary Kingston

SkipVLA divides tabletop manipulation between a classical motion planner for free-space transit and a vision-language-action (VLA) model for grasping and placing. A lightweight scorer uses 3D object boxes and frozen VLA features to select targets and trigger handoffs. In one YAM-arm task, successful MolmoAct2 trials fell from 46.56 ± 9.53 to 18.82 ± 6.27 seconds, a 59.6% reduction; findings are limited to tested pick-and-place settings.

Key insights

• Separating geometric transit from contact-rich manipulation can reduce expensive VLA inference without discarding learned visuomotor behavior.
• The hybrid design also limits the duration over which imitation-learning errors can accumulate, which may explain success-rate improvements for weaker policies such as SmolVLA.
• The approach is broadly planner-agnostic in the evaluated setting, but depends on reliable 3D detections, collision models, and discrete manipulation transitions such as grasping or releasing.

Summary by alphaXiv.