VLA / 视觉-语言-动作模型

Astra: Efficient Transformer Architecture and Contrastive Dynamics Learning for Embodied Instruction Following

Astra：面向具身指令跟随的高效Transformer架构与对比动态学习

Yueen Ma, Dafeng Chi, Shiguang Wu, Yuecheng Liu, Yuzheng Zhuang, Irwin King

机构 * Department of Computer Science and Engineering, The Chinese University of Hong Kong（计算机科学与工程系，香港中文大学） ； Huawei Noah’s Ark Lab（华为诺亚实验室） 纠错

专题命中 语言条件控制 ：vision-language-action(abstract)；action model(abstract)；分类 cs.RO

AI总结 Astra通过引入轨迹注意力和对比动态学习目标，提升了具身指令跟随任务中多模态序列处理的效率与准确性。

Comments Accepted to EMNLP 2025 (main). Published version: https://aclanthology.org/2025.emnlp-main.688/ Code available at: https://github.com/yueen-ma/Astra