摘要：We present a deep neural network to construct human-built 3D-Craft houses in Minecraft environment. Instead of hard exploration on constrained game environment, we propose a method learning to imitate human building order with recorded action sequence. Previous methods consider the action sequence as stacked voxel representation. However, the stacked voxel representation suffers from unnecessary computation cost and limited sequence information. To address these problems, we consider the action sequence as ordered point sets in 3D space. Our network is based on encoder-decoder framework. The encoder jointly learns local geometry and global sequence order information. In order to generate 3D shapes in physical environment, the decoder makes two-stream predictions, including action position and constructing block type. We conduct quantitative and qualitative experiments on 3D-Craft dataset, which demonstrates that the proposed method achieves the state-of-the-art performance in house building task.
© 2021, Springer Nature Switzerland AG.

ISSN号：0302-9743

卷、期、页：v 12572 LNCS,p541-553

发表日期：2021-01-01

收录情况：EI（工程索引）

发表期刊名称：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

参与作者：李卫民

通讯作者：刘泽宇

第一作者：刘建伟,左信

论文类型：会议论文

论文概要：刘泽宇,刘建伟,左信,李卫民，Learning 3D-Craft Generation with Predictive Action Neural Network，Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)，2021，v 12572 LNCS,p541-553

论文题目：Learning 3D-Craft Generation with Predictive Action Neural Network