面向复杂环境的机器人触觉感知算法研究综述

PDF下载HTML

Title:

Review of robot tactile perception algorithms for complex environments

作者:

高春艳, 刘琦, 李满宏, 刘璇

河北工业大学 机械工程学院, 天津 300401

Author(s):

GAO Chunyan, LIU Qi, LI Manhong, LIU Xuan

School of Mechanical Engineering, Hebei University of Technology, Tianjin 300401, China

关键词:

触觉感知; 复杂环境; 人工智能; 机器人; 深度学习; 强化学习; 多模态融合; 机器学习

Keywords:

tactile sensing; unstructured environments; artificial intelligence; robots; deep learning; reinforcement learning; machine learning; multimodal systems

分类号:

TP242.2

DOI:

10.11992/tis.202510008

摘要:

为应对复杂环境挑战并提升机器人触觉感知精度与鲁棒性，本文系统综述了近5年基于机器学习、深度学习和强化学习的典型算法，涵盖卷积神经网络(convolutional neural network，CNN)、图神经网络(graph neural network，GNN)、长短期记忆网络（long short-term memory network，LSTM）、多模态融合与主动探索。结果表明，现有方法针对噪声、非规则接触、动态时序等复杂特性已取得显著进展且各具专长，但仍面临数据稀疏、公开数据集不足、跨模态弱对齐及仿真迁移困难等共性瓶颈。结论可为低监督学习、标准数据集构建、统一多模态建模及高保真迁移研究提供参考。

Abstract:

To address the challenges posed by complex environments and improve the accuracy and robustness of robotic tactile perception, this study systematically reviews representative machine learning-, deep learning-, and reinforcement learning-based algorithms developed over the past five years, covering convolutional neural network(CNN), graph neural network(GNN), long short-term memory network(LSTM), multimodal fusion, and active exploration. The results indicate that existing methods have made significant progress in handling complex characteristics such as noise, irregular contact, and dynamic temporal patterns, while demonstrating distinct methodological advantages. Nevertheless, common bottlenecks remain, including data sparsity, limited publicly available datasets, weak cross-modal alignment, and challenges in simulation-to-real transfer. These findings provide valuable insights for future research on low-supervision learning, standardized dataset construction, unified multimodal modeling, and high-fidelity transfer in robotic tactile perception.