中国科学院自动化研究所机构知识库

微装配机器人对实现先进的微小型零件装配技术具有重要意义。令微装配机
器人具有技能学习能力，能够显著减少对人工编程开发与调试的依赖，从而进一
步提高机器人的自主性和智能性。为了实现微装配机器人的精密装配技能学习，
本文将深度学习和强化学习引入微装配机器人研究，重点开展微器件的自动识别、
抓取和装配控制等关键技术研究。论文的主要研究工作如下：
首先，为了实现微装配过程中对微零件的自动识别和高效聚焦，提出基于改
进的 SSD（Single Shot Multibox Detector）卷积神经网络的微器件自动识别方法，
并基于深度 Q 网络（Deep Q Network，DQN）建立适用于显微视觉图像的聚焦模
型，实现遮挡、光线变化和图像离焦等多种非理想成像条件下对微零件的自动识
别与快速聚焦。
其次，为了提高基于数据驱动的微零件抓取方法的探索和学习效率，设计了
一种基于深度强化学习的微零件抓取技能学习模型。机器人通过示教学习获得初
始抓取策略，然后基于好奇心机制主动探索机器人抓取环境，再利用 HER
（Hindsight Experience Replay）算法使机器人渐进地掌握微零件抓取技能。仿真
实验结果表明所提出的方法可大幅提升微零件抓取技能的探索和学习效率，并完
成机器人对微零件的抓取对准操作。
最后，为了提高微零件插入装配控制算法的适用性，提出深度强化学习和示
教学习相结合的微零件插入装配方法。在兼顾装配安全和效率的前提下，设计了
合理的微零件插入装配动作及奖励函数，机器人可实现不同外形、材质微零件的
插入装配控制。请输入中文摘要

The robot's autonomy and intelligence can be improved by implementing microassembly skill learning，which also alleviates the repeated program developing and tuning workload。 In order to realize the precision assembly skill learning of microassembly robots, this paper introduces deep learning and reinforcement learning into the research of microassembly robots, focusing on the key technologies of automatic identification, grasping and assembly control of micro-devices. The main research work of the thesis is as follows:
First of all, in order to realize the automatic identification and efficient focusing of micro parts in the micro assembly process, an automatic identification method of micro devices based on an improved SSD (Single Shot Multibox Detector) convolutional neural network is proposed, and the focusing model of microscopic vision images is established based on DQN (Deep Q Network), which can realize the automatic recognition and rapid focusing of micro parts under a variety of non-ideal imaging conditions such as occlusion, light changes and image defocusing.
Secondly, to improve the exploration and learning efficiency of data-driven micro-parts grasping methods, a microcomponent grasping skill learning model based on deep reinforcement learning is designed. The robot obtains the initial grasping strategy through teaching and learning, and then actively explores the environment based on the curiosity mechanism, and then uses the HER (Hindsight Experience Replay) algorithm to enable the robot to gradually grasp the micro part grasping skills. Experiments prove that the proposed method can greatly improve the exploration and learning efficiency of micro-parts grasping skills, and realize the grabbing and aligning operation of the microcomponent by the robot.
Finally, to improve the applicability of the micro-component insertion assembly control algorithm, a micro-component insertion assembly control method combining deep reinforcement learning and teaching learning is proposed. Under the premise of taking into account assembly safety and efficiency, a reasonable micro-part insertion assembly action and reward function are designed, and the robot can realize the autonomous decision of the optimal assembly action for micro-parts with different shapes and materials.