CASIA OpenIR  > 毕业生  > 博士学位论文
基于机理建模和数据驱动的机器人装配技能学习研究
严少华
2024-05-11
Pages148
Subtype博士
Abstract

机器人自动装配操作在航空航天、工业制造和军事等领域具有广泛的应用。随着操作环境和需求任务的复杂化,机器人需要同时掌握多种类型的装配技能,从重复进行机械运动的工具转变为具有自适应能力的智能体。但装配机器人在对准与装入过程中仍存在特征信息获取不准确、对准控制系统设计不完善、装入技能学习算法效率低且适应性差等难题。为解决上述难题,本文引入基于机理建模和数据驱动的技能学习算法,提出了针对航空接插件的基于混合视觉伺服的对准方法和基于图像视觉伺服的对准方法,以及从双轴孔装入技能到多类别轴孔装入技能的高适应性装入方法,提高了机器人对准操作的精度和速度,增强了机器人装入技能的学习效率和适应性能论文的主要工作和贡献如下:

1)针对航空接插件装配任务中位姿估计不准确和对准精度不足的问题,提出了基于U-Net目标分割方法和基于混合视觉伺服的对准控制方法。在位姿测量阶段,提出一种基于U-Net图像分割网络和椭圆特征的六维位姿测量方法。利用U-Net图像分割网络对接插件表面特征进行训练提取,并结合椭圆位姿测量算法获取航空接插件在图像空间下的部分位姿信息。然后采用基于深度结构光相机的位姿测量方案,对接插件圆环表面进行平面拟合,准确获取航空接插件在笛卡尔空间下的剩余位姿信息在对准控制阶段,利用图像空间和笛卡尔空间的不同位姿信息设计混合视觉伺服控制系统,获取机械臂末端坐标系下的运动控制量。最后,在由深度结构光相机和机械臂构成的自动装配系统中进行实验验证,结果表明所提出方法有效实现了航空接插件的高精度位姿测量与对准。

2)针对传统零件对准方法适应性能差和依赖位姿测量准确度的问题,提出了基于点、线特征和图像视觉伺服的对准方法。首先,设计了基于特征金字塔结构的U-FPN图像分割网络,利用少量标注数据训练网络,准确分割包含点、线特征的凹形区域。其次,提出了基于点、线特征交互矩阵的图像视觉伺服方法,利用双相机不同视角下的点、线特征设计位置和姿态控制律。将点特征的深度信息融合至姿态变换矩阵,并通过机械臂的主动运动对姿态变换矩阵进行标定。最后,在基于双相机的机器人自动对准装配系统中进行实验验证,通过测量不同视角下点、线特征参数,利用所提出的方法实现了零件的位姿对准与装配。与基于点特征深度估计的视觉伺服方法相比,所提出方法在对准控制精度和装配效率层面均取得了更好的效果。

3)针对大深径比的双轴孔装入任务中控制策略获取困难和受力模型建模不准确的问题,提出了基于演示学习和层次策略学习的双轴孔装入技能学习方法。首先,提出了一种基于目标的层次策略学习方法,将目标作为新变量加入到动作价值函数,以每个训练回合中达到的多个随机状态作为子目标,有效提高网络学习效率。其次,引入演示学习算法与传统控制算法相结合的初始策略,通过训练演示学习算法和传统控制算法的组合系数,提高网络动作的装入成功率和适应能力。第三,设计锥形结构的双轴孔装入仿真受力模型,模拟双轴孔装入过程中的受力情况。利用该仿真模型进行仿真训练,获取双轴孔装入技能模型。在由机械臂、力传感器组成的双轴孔装入实验系统中进行实验验证所提出方法实现了双轴孔的自动柔顺装入。该方法在训练阶段和实验阶段均取得了优于现有方法的学习效率和性能表现。

4)针对当前轴孔装入技能学习算法学习效率低、模型适应性差和训练数据缺乏的问题,提出了基于虚拟模型和自适应元策略学习的多类别轴孔装入技能学习方法。首先,设计自适应元策略学习算法,通过训练多类型轴孔装入技能模型获取适应多类型装配任务的技能。将演示学习算法构建的轨迹差异度引入状态价值函数,并采用反馈控制算法作为初始策略,提升算法的学习性能。其次,分别在仿真软件和受力分析模型中构建单、双、三轴孔装入仿真环境,利用仿真环境进行训练并获取技能模型。在由机械臂、力传感器、夹爪组成的多类别轴孔装入实验平台中进行实验验证,所提出方法实现了从单、双、三轴孔至多孔插座的自适应柔顺装入。相对于现有的技能学习方法,该方法取得了更好的装入效率、装入性能和适应能力

最后,对本文工作和研究成果进行了总结,并对未来的研究工作进行了分析与展望。

Other Abstract

Robotic automated assembly operations have a wide range of applications in aerospace, industrial manufacturing, and military. With the complexity of the robot operating environment and demanded tasks, robots need to master multiple types of assembly skills, transforming from tools that perform repetitive mechanical movements to intelligent agent with adaptive capabilities. However, there are a series of difficulties such as inaccurate acquisition of feature information, imperfect design of alignment control system, low learning efficiency and poor adaptability of insertion skill learning algorithms for assembly robots in the process of alignment and insertion. In this dissertation, the skill learning methods based on mechanism modeling and data driven are introduced to improve the accuracy and speed of alignment operations, and enhance the learning efficiency and adaptive performance of insertion skills. A series of methods of skill learning for robot assembly are proposed, including the hybrid visual servoing alignment method and the image-based visual servoing alignment method for aerial connectors, and the highly adaptive insertion method from dual peg-in-hole insertion skill to multi-category peg-in-hole insertion skill. The main work and contributions of this dissertation are as follows:

(1) Aiming at the problems of inaccurate position estimation and low alignment accuracy in aviation connector assembly tasks, a U-Net target segmentation method and a hybrid visual servoing alignment control method are proposed. In the stage of position measurement, a six-dimensional pose measurement method based on U-Net image segmentation network and elliptical features is proposed. The U-Net image segmentation network is utilized to train and extract the surface features of the connector, and combined with the elliptical position measurement algorithm to obtain the partial position information of the aerial connector under the image space. A position measurement scheme based on depth structured light camera is used to fit the plane with the surface of the connector ring and accurately obtain the remaining pose information of the aerial connector in Cartesian space. In the alignment control stage, the different pose information under image space and Cartesian space is utilized to design a hybrid visual servoing control system to obtain the motion control quantity under the end-effector coordinate system of manipulator. The proposed method is verified by experiments in an automatic assembly system consisting of a deep structured light camera and a manipulator. The proposed method can effectively realize the high-precision pose measurement and alignment of the aerial connector.

(2) Aiming at the problems of poor adaptive performance and dependence on the accuracy of position measurement of traditional alignment methods, a alignment method based on point and line features and image-based visual servoing is proposed. Firstly, a U-FPN image segmentation network based on the feature pyramid structure is designed, and a small amount of labeled data is used to train the network. The network model can accurately segment concave regions containing point and line features. Secondly, an image-based visual servoing method based on the interaction matrixes of point and line features is proposed, and the position and orientation control laws are designed by the point and line features under different views of two cameras. The depth information of the point features is fused to the pose transformation matrix, and the pose transformation matrix is calibrated by the active motion of the manipulator. Finally, the proposed method is experimentally verified in a dual-camera-based robotic automatic alignment assembly system. The pose alignment and assembly of parts are achieved using the proposed method based on point and line features under different views. Compared with the visual servoing method based on the depth estimation of point features, the proposed method achieves better results in both alignment control accuracy and assembly efficiency.

(3) Aiming at the problems of difficulty in control policy acquisition and inaccurate force model modeling in dual peg-in-hole insertion tasks with large depth-to-diameter ratios, a dual peg-in-hole insertion skill learning method based on demonstration learning and hierarchical policy learning is proposed. Firstly, a goal-based hierarchical policy learning method is proposed, which adds the goal as a new variable to the action value function, and uses multiple random states reached in each training round as subgoals to effectively improve the network learning efficiency. Secondly, the initial policy of combining the demonstration learning algorithm with the traditional control algorithm is introduced to improve the insertion success rate and adaptive ability of network actions by training the combined coefficients of the demonstration learning algorithm and the traditional control algorithm. Thirdly, the simulation force model of dual peg-in-hole insertion with conical structure is designed to simulate the force during dual peg-in-hole insertion. The simulation model is utilized for simulation training to obtain the dual peg-in-hole insertion skill model. Experiments are conducted with the dual peg-in-hole insertion experimental system consisting of a manipulator, a force sensor, and the effectiveness of the dual peg-in-hole compliant insertion control method is verified. The proposed method achieves better learning efficiency and performance than the existing methods in both the training phase and the experimental phase.

(4) Aiming at the problems of low learning efficiency, poor model adaptability and lack of training data of the current peg-in-hole insertion skill learning algorithm, a multi-category peg-in-hole insertion skill learning method based on virtual model and adaptive meta policy learning is proposed. Firstly, the adaptive meta policy learning algorithm is designed to train skill models simultaneously to obtain multi-category peg-in-hole insertion skills adapted to multi-category assembly tasks. The trajectory difference function constructed by the demonstration learning algorithm is introduced into the state value function, while the feedback control algorithm is adopted as the initial policy to improve the algorithm learning performance. Secondly, single, dual and triple peg-in-hole insertion simulation environments are constructed in simulation software and force analysis model, respectively. The simulation environment is utilized for the proposed algorithm to train the skill models. A multi-category peg-in-hole insertion experimental platform consisting of a manipulator, a force sensor, and a gripper is constructed. Insertion experiments are completed and the effectiveness of the adaptive and compliant insertion control method is verified for single, dual, and triple peg-in-hole and multi-hole sockets insertion tasks. The proposed method achieves better insertion efficiency, insertion performance and adaptive capability compared to existing skill learning methods.

Finally, the research results of this dissertation are summarized, and future research is analyzed and prospected.

Keyword位姿测量 对准控制 强化学习 装入控制 装配技能学习
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56676
Collection毕业生_博士学位论文
Corresponding Author严少华
Recommended Citation
GB/T 7714
严少华. 基于机理建模和数据驱动的机器人装配技能学习研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
基于机理建模和数据驱动的机器人装配技能学(6690KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[严少华]'s Articles
Baidu academic
Similar articles in Baidu academic
[严少华]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[严少华]'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.