具有自主发育能力的机器人感知与认知方法研究

发布时间：2019-05-24 16:30

【摘要】：集装箱装卸自动化是运输集装箱化的必然要求,在当前集装箱装卸作业中,扭锁的装卸仍然由人工来完成,这不仅增加了劳动强度,降低了生产效率,还严重威胁到工人的人身安全,亟需以机器人为核心的自动化技术来取代人工操作。本文以海港集装箱扭锁的自动化安装为研究背景,根据扭锁安装需求搭建模拟平台,主要解决扭锁的认知识别与抓取位姿估计问题。由于扭锁种类繁多,且随着需求不断改进更新,抓取任务不断有新的挑战,机器人的认知系统需要在线实时地更新、存储新的特征,否则无法准确识别新类别物体。而传统机器人的认知系统存在任务确定、离线学习、实时性差及自适应性差等问题,无法完成非特定任务。为了解决从工作场景中识别并准确抓取指定物体的问题,针对传统机器人认知系统存在的局限性,从认知机器人的研究思路出发,模拟人类学习方式、智能表现形式以及人脑智能信息处理机制,建立本文的机器人认知系统,使机器人通过在线学习,将累积的知识和经验动态有组织地存储到记忆系统中,在执行任务时回调以往的经验知识做出准确的识别,进而获取准确的位姿估计。基于自主发育范式将扭锁抓取机器人的认知系统分为:感知发育、认知发育以及任务执行三大模块,从三个功能模块展开,本文主要的研究工作如下:(1)传感器数据预处理,本文提出了基于分域策略的联合双边滤波预处理方法,解决了Kinect传感器采集的深度图像存在漏洞、不对齐以及噪声等问题。根据Kinect传感器三种误差来源的区域特性,对深度图像进行分区域滤波处理。根据深度图像和彩色图像的结构相关性,对深度像素进行分类,将漏点及不对齐像素归类为不可信任区域,其余像素归为可信任区域。融合彩色图像信息,采用联合滤波方法对深度图像进行引导滤波,针对可信任区域像素采用联合三边滤波方法;针对不可信任区域中边缘像素采用Sigmoid-方向高斯的联合双边滤波方法,非边缘像素采用Sigmoid-颜色相似的联合双边滤波方法。其中,基于增强学习中的奖惩原则,使用Sigmoid函数为不可信任区域像素动态产生置信度空域权重,赋予滤波邻域内与中心点属性相同的可信任信息较高权重;使用方向高斯滤波函数为边缘像素产生颜色权重,赋予滤波邻域内与边缘方向一致的像素较高权重,保留边界方向性;基于可信度势场理念选取滤波方向,确保滤波邻域内含有更多有效的与待滤波点属性相同的可信任信息,通过以上策略手段来保证滤波后深度信息的合理性和准确性。最后通过对比实验从视觉度量、降噪性能及运行时间上,有力地证明了本文滤波方法的优越性能。(2)本文提出了在线自适应增量PCA学习方法,解决了感知发育中特征提取与数据降维问题。该方法能够在线自主地发现和选择输入数据的有效特征,更新优化特征空间,发育出适合机器人内部表达的模型。针对PCA学习方法对样本数量及多样性依赖程度高、缺乏自适应性、不能在线增量更新、可扩展性差等问题;增量PCA方法随着样本输入,特征维度、计算量和存储量都随之增加等问题。本文算法在增量PCA的基础上进行改进,基于新样本与已有特征空间重建样本之间的差异程度监测新类别输入,控制特征空间增量地更新;基于类内距离比较,自适应地更新类内距离阈值,优化特征空间向量。实验表明该算法在少量训练样本的情况下,能够在线地学习、更新与优化、累积新特征,将高维输入信号合理降维,增强了视觉系统的感知和识别能力。(3)本文借鉴人脑记忆系统中前额叶、海马以及海马前额叶回路的信息处理机制,提出了三层的基于增量式神经网络的认知发育模型,能够在线对所学的知识和经验实时有效地存储、累积、整合以及回调,解决传统数据库存储知识的固定性、封闭性等问题,更好地适应未知的动态环境。认知发育网络中有监督学习和无监督学习方式可同时并存,随着与外界不断的交互,中间层神经元同时接受外界通过效应层传递的自上而下的监督指导信号和来自输入自底向上的响应信号,使用Hebbian学习规则来模拟神经元学习响应过程,采用Top-K竞争机制模拟神经元的侧抑制效应,引入遗忘平均函数产生权重模拟人类接受新知识的速度,通过以上策略模拟大脑皮层理解、记忆情况。认知发育神经网络在第四章感知发育模块基础上,基于重建误差控制神经网络节点的增加,基于熟悉相似度控制被激活神经元的权重更新。通过实验表明,认识发育网络可以将学习的结果以“知识”的形式有组织地、动态地存储到记忆系统中,取代传统数据库,提高了扭锁的准确识别率。(4)扭锁抓取位姿估计,本文根据扭锁安装需求搭建抓取平台,经认知分析后获取扭锁正确类别及其正反面信息,与相应类型的标准位姿做比对,将位姿估计问题简化为两个点云集匹配问题,采用迭代最近点(ICP)算法估算可抓取点的位置和姿态,为下一步抓取规划提供数据支持。通过实验,证明了该方法的可行性。最后,总结全文所做的工作,提出今后进一步需要研究的问题。
[Abstract]:The container loading and unloading automation is an inevitable requirement for the transportation of the container. In the present container loading and unloading operation, the loading and unloading of the twist lock is still carried out manually, which not only increases the labor intensity, reduces the production efficiency, but also seriously threatens the personal safety of the workers, Robotic-based automation technology is needed to replace manual operations. The paper takes the automatic installation of the twist lock of the harbor container as the research background, and sets up the simulation platform according to the installation requirements of the twist lock, and mainly solves the problem of the cognition recognition and the grasping pose estimation of the twist lock. Due to the wide variety of twist locks, and with the continuous improvement of the demand, the grasping task has new challenges, and the robot's cognitive system needs to be updated online in real time, and new features can be stored, otherwise, the new category object cannot be accurately identified. The cognitive system of the traditional robot has the problems of task determination, off-line learning, poor real-time performance and poor self-adaptability. In order to solve the problem of identifying and accurately capturing the specified object from the work scene, aiming at the limitation of the traditional robot cognitive system, the human learning method, the intelligent expression form and the human brain intelligent information processing mechanism are simulated from the research thinking of the cognitive robot, In this paper, the robot cognitive system is established, which enables the robot to dynamically organize the accumulated knowledge and experience into the memory system through on-line learning, and to make an accurate identification of the past experience knowledge when executing the task, so as to obtain the accurate pose estimation. The cognitive system of the twist-lock grasping robot is divided into three modules: the sense development, the cognitive development and the task execution based on the independent development paradigm, and the main research work in this paper is as follows: (1) the sensor data is pre-processed, In this paper, a combined double-side filtering pre-processing method based on the split-domain strategy is proposed, and the problems such as the vulnerability, the misalignment and the noise of the depth image acquired by the Kinect sensor are solved. According to the region characteristics of three error sources of the Kinect sensor, the depth image is divided into region filtering processing. According to the structure correlation of the depth image and the color image, the depth pixel is classified, and the missing point and the non-aligned pixel are classified as the non-trusted area, and the remaining pixels are classified as a trusted area. the method comprises the following steps of: fusing the color image information, carrying out direct filtering on the depth image by using a joint filtering method, adopting a combined trilateral filtering method for the trusted area pixels, and adopting a joint bilateral filtering method of the sigmoid-direction gauss aiming at the edge pixels in the non-trusted area, The non-edge pixels adopt a joint double-sided filtering method similar to the Simoid-color. The method comprises the following steps of: dynamically generating a confidence spatial weight for a non-trusted area pixel by using a Simoid function based on the reward and punishment principle in the enhanced learning, and giving a high weight of the trusted information which is the same as the center point attribute in the filter neighborhood; and generating a color weight for the edge pixel by using the directional Gaussian filter function, and the filtering direction is selected based on the concept of the reliability potential field to ensure that more effective trust information is contained in the filter neighborhood which is the same as that of the point to be filtered, And the rationality and the accuracy of the depth information after filtering are ensured through the above strategy means. Finally, the superiority of the filtering method in this paper is proved by the contrast experiment from the visual measurement, the noise reduction performance and the running time. (2) In this paper, an on-line self-adaptive incremental PCA learning method is proposed to solve the problem of feature extraction and data reduction in sensing development. The method can automatically discover and select the effective characteristics of the input data, update the optimized feature space, and develop a model suitable for the internal expression of the robot. The method of PCA learning has the problems of high sample number and diversity, lack of self-adaptability, no on-line incremental updating, poor scalability, etc. The increment PCA method increases with the sample input, the feature dimension, the calculation quantity and the storage amount. the algorithm is improved on the basis of the increment PCA, the new category input is monitored based on the difference between the new sample and the existing feature space reconstruction sample, the control feature space is updated incrementally, the intra-class distance threshold is adaptively updated based on the intra-class distance comparison, The feature space vector is optimized. The experiment shows that the algorithm can study, update and optimize on-line, accumulate new features in a small amount of training samples, reduce the dimension of the high-dimension input signal, and enhance the perception and recognition ability of the vision system. (3) Based on the information processing mechanism of the prefrontal lobe, the hippocampus and the frontal lobe of the hippocampus of the brain memory system, a three-layer cognitive development model based on the incremental neural network is proposed, which can effectively store and accumulate the learned knowledge and experience in real time. And the problem that the traditional database storage knowledge is fixed, closed and the like is solved, and the unknown dynamic environment is better adapted. in that cognitive development network, the supervised learning and the non-supervised learning method can coexist at the same time, and as the interaction with the external environment, the middle-layer neuron receives the top-down supervision guidance signal transmitted by the outside through the effect layer and the response signal from the input self-bottom, Using the Hebbian learning rule to simulate the learning response of the neuron, the side effect of the neuron was simulated by the Top-K competition mechanism, and the forgetting average function was introduced to generate the weight to simulate the speed of the human being's new knowledge. The above strategy was used to simulate the understanding and memory of the cerebral cortex. The cognitive development neural network, based on the fourth-sense development module, controls the increase of the neural network node based on the reconstruction error, and controls the weight update of the activated neuron based on the familiar similarity. The experiment shows that the cognitive development network can be organized and dynamically stored in the memory system in the form of "knowledge", instead of the traditional database, the accurate recognition rate of the twist lock is improved. (4) the position and position estimation of the twist lock is constructed, a grasping platform is built according to the installation requirement of the twist lock, the correct category of the twist lock and the positive and negative information of the twist lock are acquired through the cognitive analysis, and the pose estimation problem is simplified into two point cloud matching problems, An iterative recent point (ICP) algorithm is used to estimate the position and attitude of the grab points and provide data support for next-step grab planning. The feasibility of this method is proved by the experiment. Finally, the paper sums up the work done in the whole text, and puts forward some problems that need to be studied in the future.
【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41;TP242

【参考文献】