• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 257-268.doi: 10.3901/JME.260729

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

基于关系推理和主动训练的室内语义地图预测方法

张嘉东, 王巍   

  1. 北京航空航天大学机械工程及自动化学院 北京 100191
  • 收稿日期:2025-12-31 修回日期:2026-06-01 发布日期:2026-11-20
  • 作者简介:张嘉东,男,1996年出生,博士研究生。主要研究方向为语义建图、物体建模。E-mail:zhangjiadong@buaa.edu.cn;王巍(通信作者),男,1972年出生,博士,教授,博士研究生导师。主要研究方向为机器人定位与建图、机器人控制与规划。E-mail:wangweilab@buaa.edu.cn
  • 基金资助:
    中央高校基本科研业务费(JK2024-50-02)和内蒙古自治区科技计划产业技术创新资助项目。

Indoor Semantic Map Prediction Based on Relationship Reasoning and Active Training

ZHANG Jiadong, WANG Wei   

  1. School of Mechanical Engineering and Automation, Beihang University, Beijing 100191
  • Received:2025-12-31 Revised:2026-06-01 Published:2026-11-20

摘要: 精确且完整的语义地图对于物体搜索至关重要。然而在真实室内环境中,由于物体遮挡、机器人视野受限、空间布局多样化等因素,语义地图预测与物体搜索的性能被严重限制。针对上述问题,提出一种基于关系推理和主动训练的语义地图预测方法。首先,构建栅格级语义空间关系推理模块,通过语义关系图和空间关系图建模物体类别共现规律与空间邻近关系,从而增强未观测区域的语义补全能力。其次,提出基于信息增益的主动训练策略,利用集成模型估计各栅格的认知不确定性,引导机器人主动采集高价值训练样本,以提升模型在未知环境中的泛化能力。最后,设计不确定性驱动的物体搜索机制,通过置信上界评分函数联合考虑目标类别预测概率与不确定性,实现探索与利用的动态平衡。在基于Matterport3D数据集的仿真室内环境中对所提方法进行了验证,实验结果表明在语义建图精度与物体搜索性能方面均优于对比方法。

关键词: 语义建图, 地图预测, 物体搜索, 关系建模

Abstract: An accurate and complete semantic map is crucial for effective object search. However, in real indoor environments, factors such as object occlusion, limited field of view, and diverse spatial structures can significantly degrade the performance of semantic prediction and object search. To address these challenges, a semantic map prediction method based on relationship reasoning and active training is proposed. First, a grid-level semantic–spatial reasoning module is developed to model category co-occurrence and spatial proximity through semantic and spatial relation graphs, enhancing semantic completion in unobserved regions. Second, an information-gain-based active training strategy estimates grid-wise epistemic uncertainty using an ensemble model to guide the robot toward informative samples. Finally, an uncertainty-driven search mechanism uses an upper confidence bound score to jointly consider target probability and uncertainty, balancing exploration and exploitation. The proposed method is evaluated in simulated environments based on the Matterport3D dataset and demonstrates significant improvements in both semantic mapping and object search performance.

Key words: semantic mapping, map prediction, object search, relationship modeling

中图分类号: