VLA ModelsarXiv preprint2025

GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization

Yi-Ting Hsiao, Siavash Khodadadeh, Kevin Duarte, Wei-An Lin, Hui Qu, Mingi Kwon, Ratheesh Kalarot et al.Stanford University

摘要

We present GuidedVLA, a plug-and-play approach that specializes VLA action generation through attention mechanisms guided by task-relevant factors. Our method allows users to specify constraints and preferences without retraining, enabling more precise and controllable robot manipulation.

VLAattention specializationtask specificationplug-and-playrobot manipulation

技术细节

仿真平台
测试机器人
模型骨架

VLA Transformer

编码器

SigLIP

解码器

Attention-specialized Action Head

相关公司(1)

京ICP备2026064258号-1