CASIA OpenIR  > 精密感知与控制研究中心
Thesis Advisor张大朋
Degree Grantor中科院自动化所
Place of Conferral智能化大厦
Degree Discipline控制工程
Keyword演示示教 域自适应 轨迹模仿 深度强化学习



       二、针对常规微零件示教趋近轨迹的起点和终点位置固定的问题,本文提出基于动态运动基元的微零件趋近轨迹模仿学习方法,可实现微零件起始点变化时的趋近轨迹的自主生成。本文首先采用虚拟现实设备采集微零件趋近示教数据,并通过 DTW 算法进行数据对齐;然后采用 GMM 和 GMR 算法学习并生成最优示教轨迹;最后通过 DMP 模型学习并泛化微零件趋近轨迹,实现微零件起始位置变化时趋近轨迹的自主生成。
       三、针对过渡配合的微零件装配时无法兼顾装配力和位置的问题,本文首先建立基于深度强化学习的机器人装配控制模型,并通过演示示教数据缩短模型训练时间;然后采用连续性控制策略 TD3 算法确定当前装配状态下的最优装配动作,实现兼顾微零件装配力和位置的应用目标。

Other Abstract

       The micro-assembly robots with automatic identification and decision-making capabilities can significantly reduce the dependence on manual programming and tuning, and effectively improve the assembly effciency of micro-parts. This thesis introduces domain adaptation, dynamic primitives and deep reinforcement learning into the research of micro-assembly robots in order to achieve the precision assembly skill learning, focusing on the key technologies of micro-part recognition and localization, trajectory imitation, and force/position assembly control. The main contents of the thesis are as follows:

       1.Considering the high cost of manual part annotation and poor model adaptability, this thesis proposes an automatic identification method based on domain adaptation to localize micro-parts under different textures. First, a micro-part synthetic semantic segmentation dataset is proposed, which automatically generate micro-part image annotations with the help of computer graphics. Then the style transfer method is applied to generate micro-part images with different texture backgrounds to improve the diversity of micro-part training data. Finally, semantic segmentation is adopted to extract the feature of micro-parts through pixel-wise domain adaptation.

     2.In this paper, the micro-part approaching trajectory model is proposed based on dynamic motion primitives, which can realize automatic approaching of microparts under the condition of changing starting points. First, virtual reality equipment is used to collect the approaching teaching data of micro-parts. DTW algorithm is applied for time alignment. Then GMM and GMR algorithms are utilized to learn and generate the optimal teaching trajectory. Finally, the approaching trajectory of micro-parts is generated to achieve the goal of the micro-parts, which can automatically approach the assembly station with the starting position changing.

     3.Since the assembly force and position cannot be considered in the assembly of micro-parts, this thesis first establishes robot assembly control model based on deep reinforcement learning, and shortens the training time of the control model by finetuning with the teaching data. Then, the continuity control strategy TD3 method is used to determine the optimal assembly action in the current assembly state, taking both the assembly force and position of the micro-parts into consideration.


Document Type学位论文
Recommended Citation
GB/T 7714
夏鹏程. 基于演示示教的机器人技能模仿学习[D]. 智能化大厦. 中科院自动化所,2021.
Files in This Item:
File Name/Size DocType Version Access License
夏鹏程-学位论文.pdf(16930KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[夏鹏程]'s Articles
Baidu academic
Similar articles in Baidu academic
[夏鹏程]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[夏鹏程]'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.