【作者】 王浩；

【导师】 彭刚；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2022， 硕士

【摘要】 目前机器人工业应用大多依赖于结构化的生产环境,通过人工示教的方法完成一些特定动作,而近年来仓储物流、家庭环境等非结构化环境的需求日益增多,机器人技术的智能化成为主流发展方向。机器人对接与抓取任务尤为常见与重要,是当前机器人领域的研究热点,深入研究机器人对接与抓取技术对推进机器人智能化进程和提高生产效率具有重要意义。本文面向机械臂任务中常见的对接与抓取任务,研究基于深度视觉的机械臂对接与抓取规划方法。针对加油对接任务中加油口特征不明显问题,提出了一种基于RGB-D的加油对接规划方法,利用RGB检测定位加油口,深度点云计算加油口位姿,并通过调整视角提高鲁棒性。同时,为解决因反光造成的点云缺失问题,进一步提出基于三维重建的加油对接规划方法,利用三维重建补全加油口点云数据,提高了对接精度。在6-DOF抓取任务中,提出了一种基于深度学习的6-DOF抓取规划框架,包含单阶段、双阶段两种方法,分别通过抓取姿态预测网络,以及抓取姿态预测与评估网络,获取高质量抓取姿态,提高了抓取成功率与实时性。为进一步提高真实环境抓取规划鲁棒性,在双阶段抓取规划方法的基础上,针对真实抓取数据集难获取问题,提出了一种结合自监督技术的6-DOF抓取规划方法,控制机械臂采用推动策略采集多视角数据,通过三维重建获取场景完整点云,进行数据标注,获得真实抓取数据集。同时,为了提高网络预测性能,引入复用结构到抓取姿态预测与评估网络中,并增加位姿计算模块,进一步提高抓取姿态精度。本文搭建了基于深度视觉的机械臂对接与抓取实验系统,实验结果表明,所提出的机械臂规划方法可以较好地完成对接与抓取任务,提高对接和抓取的精度或成功率,并具有较好的鲁棒性和实时性。更多还原

【Abstract】 At present,industrial application of robots mostly relies on structured production environment,in which some specified actions are completed by manual demonstration method.In recent years,the demand in some unstructured environments such as warehousing and logistics and home environments is increasing and intelligence of robotics becomes the mainstream development direction.Robot docking and grasping tasks are particularly common and important,which are the current research hotspots in the field of robotics.An in-depth study on robot docking and grasping technology is of great significant to advance robot intelligence progress and improve production efficiency.In this thesis,robot arm docking and grasping planning method based on a deep vision for the common docking and grasping tasks in robot arm tasks.Aiming to the problem of inconspicuous features of refueling port in refueling docking tasks,the thesis proposes a refueling docking planning method based on RGB-D that refueling port is located by using RGB detection,the position and orientation of the refueling port is calculated through deep point cloud,and robustness is improved by adjusting the viewpoint.Meanwhile,to solve the problem of missing point clods caused by reflections,the thesis proposes refueling docking planning method based on threedimension reconstruction that docking accuracy is improved by using three-dimension reconstruction to complete point cloud data of refueling port.In the 6-DOF grasping task,the thesis proposes a 6-DOF grasping planning framework based on deep learning,which includes two methods,single-stage and two-stage.The two methods obtain high-quality grasping posture by using grasping posture prediction network and grasping posture prediction and evaluation network respectively,which improves the success rate of grasping and real-time performance.To further improve the robustness of real environment grasp planning,based on the two-stage grasp planning method,a 6-DOF grasping planning method combined with self-supervision technology is proposed for difficulty in obtaining real grasp data sets,where robot arm is controlled to adopt a push strategy to collect multiview data,obtain the complete point cloud of scene through 3D reconstruction,and perform data annotation to obtain real grasp data sets.Meanwhile,in order to improve the network prediction performance,a multiplexing structure is introduced into the grasping posture prediction and evaluation network,and position and orientation calculation module is added to further improve the accuracy of grasping posture.In this thesis,a robot arm docking and grasping experimental system based on deep vision is built,and the result shows that the proposed robot arm planning method can better complete the docking and grasping tasks,improve the accuracy or success rate of docking and grasping,and have better robustness and real-time performance.更多还原