摘要
针对机器人在双臂协同插孔等精细操作中运动控制策略训练复杂、路径规划执行不够高效和精确的问题,提出了一种基于多尺度特征金字塔和注意力机制的改进机器人行为克隆算法。结合残差网络和特征金字塔设计backbone网络以提取并融合多尺度图像特征,加强机器人的环境感知和视觉反馈能力;引入动作分块模块以提升控制策略的准确和平滑,降低行为克隆中的复合误差;基于注意力机制将控制策略训练成为条件变分自编码器(CVAE)以学习示教数据分布和捕获图像特征与动作的相关性,从而提高策略的泛化能力和面对陌生环境的适应性及创造性。仿真结果表明,所提算法在两个精细操作任务中的成功率和轨迹平滑显著优于其他五个基线模型,充分证明了该算法能够通过简单训练精确地执行机器人精细操作任务。
This paper addressed the challenges of complex motion control strategy training,inefficient and imprecise path planning execution in robots performing fine operations such as dual-arm cooperative insertion tasks.It proposed an improved robot behavior cloning algorithm based on multiscale feature pyramids and attention mechanisms.The algorithm combined residual networks and feature pyramids to design the backbone network,extracting and fusing multi-scale image features,which enhanced the robot’s environmental perception and visual feedback capabilities.It introduced an action segmentation module to improve the accuracy and smoothness of control strategies,reducing compound errors in behavior cloning.Additionally,the algorithm trained the control strategy as a conditional variational autoencoder(CVAE)using the attention mechanism to learn the distribution of demonstration data and capture the correlation between image features and actions.This approach improved the generalization ability and adaptability of the strategy in unfamiliar environments.Simulation results show that the proposed algorithm outperforms five baseline models in terms of success rate and trajectory smoothness in two fine operation tasks.These results demonstrate that the algorithm can execute precise robot fine operation tasks through simple training.
作者
黄小霞
阳波
向鑫
陈灵
陈中祥
孙舜尧
肖宏峰
Huang Xiaoxia;Yang Bo;Xiang Xin;Chen Ling;Chen Zhongxiang;Sun Shunyao;XiaoHongfeng(College of Engineering&Design,Hunan Normal University,Changsha 410081,China)
出处
《计算机应用研究》
北大核心
2025年第6期1668-1675,共8页
Application Research of Computers
基金
国家自然科学基金资助项目(12271525)
湖南省科技创新计划资助项目(2024JK2022)。
关键词
机器人精细操作
运动控制策略
行为克隆
动作序列
precision operation of robot
motion control strategy
behavioral cloning
action sequence
