SIA OpenIR  > 机器人学研究室
基于学习方法的机器人轴孔装配问题研究
Alternative TitleResearch on the Robotic Peg-in-Hole Assembly based on Machine Learning
刘乃龙
Department机器人学研究室
Thesis Advisor王志东 ; 崔龙
Keyword机器人装配 力控制 示教学习 深度强化学习 接触状态
Pages110页
Degree Discipline机械电子工程
Degree Name博士
2020-05-22
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract装配操作广泛存在于工业制造和生产过程中。由于机器人灵巧操作能力的限制,至今仍有大量的装配任务无法被机器人所取代。传统的机器人的操控方法往往需要精确的本体模型和环境模型,而在装配操作中机器人与环境接触的动力学难以被精确建模,特别是与环境之间的摩擦力模型。在接触任务中存在着诸多不确定性,导致传统的方法在解决机器人装配问题中具有很大的局限性。随着机器学习方法的快速发展,特别是它模式识别领域,尤其是计算机视觉领域有了快速的发展,而被认为是一种非常有前途的方法。机器学习方法不需要准确指定系统模型,而是通过数据驱动方法来建模。虽然机器学习方法在模式识别领域有了很大的发展,但是与机器人操作控制的结合还存在诸多限制。本文试图通过机器学习技术,与机器人传统的操控技术相结合,以提高机器人的操作能力,来使得机器人能够学习复杂的装配操作技巧,而无需建立精确的环境接触模型。 传统的机器人任务往往依赖大量的手动编程和硬编码的方法,难以适应这种快速的产品线变更。针对传统方法的这种局限性,我们提出在 3D 图形仿真环境下基于深度强化学习的自学习方法来训练机器人学习轴孔装配任务,利用深度神经网络来编码机器人控制策略,通过 PPO 深度强化学习方法来训练该深度神经网络,以泛化机器人的操作策略,使得机器人能处理更多环境不确定性因素下的轴孔装配问题。在基于 ROS 和 Gazebo 搭建的仿真平台上,进行仿真实验,证明我们提出的方法的有效性。 示教学习方法被认为是解决机器人进行复杂操作任务的一种很有前途的方法,但是以传统的示教学习方法是为了学习位置轨迹,并没有考虑环境的接触力,为此我们基于传统的笛卡尔空间 DMPs 轨迹生成模型提出了一种改进的笛卡尔空间 DMPs 模型,使其生成的轨迹更接近于人类操作者的运动行为。同时为了学习运动过程中的接触行为,我们在模型中引入了接触力和位置的反馈机制,使得模型模型不仅可以编码位置轨迹信息,同时也能编码接触行为中的力信息。同时,在执行轨迹的装配过程中,我们使用了基于阻抗控制的柔顺控制器,以保证安全柔顺的操作行为。并通过仿真实验验证了轨迹生成模型的收敛性,并在真实硬件上说明了改进后的这种带有力位混合反馈机制的模型在插孔过程中的动态性能有了一定的提升。 进行接触状态的辨识和分析,有助于进一步设计更有针对性的控制系统,同时也有助于对状态过程的装配事件进行监控。传统的接触状态辨识的方法往往都是通过监督数据的方式对装配过程中所产生的接触信号特征和接触状态之间建立映射关系,从而对接触状态进行分类。这种方法的局限性就是依赖于这种监督数据的获取,另外其往往都依赖于准静态接触模型。为了进一步发现更多隐藏的接触类型,我们直接从原生的接触信号出发,利用无监督和多变量时间序列的聚类方式,从接触数据中获取更多的接触类型。通过利用 DTC 网络构建端到端的接触状态聚类模型,来对装配过程中的状态进行辨识。 大部分的机器学习方法需要大量的数据来进行长时间的训练,而在真实环境中,这种成本是很高的,为此,我们搭建了数据更容易产生和获取的 3D 仿真环境。并针对示教学习方法,利用 KUKA 机器人搭建了学习平台。最后基于该平台进行仿真的实验研究,以表明本文相关研究方法的有效性。
Other AbstractAssembly manipulations are widely found in industrial manufacturing and production processes. Due to the limitation of the robot’s dexterous manipulation ability, there are still a large number of assembly tasks that cannot be replaced by robots. Traditional robot manipulation and control methods often require an accurate dynamics model about the robot and environment, while the contact dynamics of assembly manipulation are difficult to accurately model, especially frictional models between the environment. There are many uncertainties in contact tasks, which lead to the traditional methods in solving the problem of robot assembly has great limitations. With the rapid development of machine learning methods in the field of pattern recognition, especially in the field of computer vision, it is considered to be a promising method. Machine learning methods do not need to specify system models accurately, but instead, model them through data-driven methods. Although machine learning methods have developed significantly in pattern recognition, there are still many limitations in the combination of robot manipulation and control. This paper attempts to improve the robot's ability to operate by combining machine learning technology with the robot's traditional manipulation techniques. That way, the robot can learn complex assembly skills without the need to establish a precise model of environmental contact. Traditional robotic tasks often rely on manual programming and hard-coding methods, making it difficult to adapt to such rapid product line changes. For this limitation of traditional methods, we propose to train robots to learn peg-in-hole tasks based on deep reinforcement learning in the 3D graphics simulation environment, to encode robot control strategies by using deep neural networks, to train the deep neural network through PPO algorithm, then generalize the robot's operating strategy, so that the robot can deal with the problem of PiH assembly under more uncertainties. Simulation experiments are conducted on a simulation platform based on ROS and Gazebo to demonstrate the effectiveness of the proposed method. Learning from demonstration(LfD) is considered to be a promising method to solve the complex manipulation task of robots. However, the traditional LfD method is to learn the position trajectory. It does not consider the contact force of the environment. For this reason, we propose an improved Cartesian space DMPs model based on the traditional Cartesian space DMPs trajectory model, so that the resulting trajectory is closer to the human operator’s motor behavior. In order to learn the contact behavior in the motion, the feedback mechanism of contact force and position is introduced into the model so that it can encode not only the position track information but also the force information in the contact behavior. During the assembly process, we used a compliance controller based on impedance control to ensure safe and smooth operation. And the convergence of the trajectory generation model is verified by simulation experiments, and the improved model with force and position feedback mechanism is verified on the real hardware. The dynamic performance of the assembly process can be improved by the model with a hybrid force-position feedback mechanism. Identification and analysis of contact state can help to design more targeted controller further, as well as to monitor assembly events in the process. The traditional method of contact state identification is often to classify the contact state by the supervised dataset to establish a mapping relationship between the contact signal characteristics and contact state produced during assembly. The limitation of this approach is that it relies on the acquisition of supervised data, and it often relies on quasi-static contact models. To discover more hidden contact types, we obtain time series from contact data by clustering them from unsupervised and multivariate time series with the raw contact signals. The contact state is identified by building an end-to-end contact state clustering model using the DTC network. Most machine learning methods require much data for long training, which is expensive in real-world environments, so a 3D simulation environment is built where data is more easily generated and acquired. For using the LfD methods, the KUKA robot is used as a learning platform. Finally, the experimental research based on the platform is carried out to show the validity of the relevant research methods in this paper.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27164
Collection机器人学研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
刘乃龙. 基于学习方法的机器人轴孔装配问题研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
基于学习方法的机器人轴孔装配问题研究.p(12827KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘乃龙]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘乃龙]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘乃龙]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.