毕业生知识库

随着机器人应用从传统的工业领域向服务领域的不断扩展，新的环境与任务对机械臂的灵活抓取功能提出了新的要求。因此，如何在复杂堆叠环境中实现对目标物体灵巧的抓取仍是当前研究的热点和难点。本文研究非结构化场景下日常物品的抓取学习方法，从6自由度抓取位姿检测、领域自适应特征学习和机器人抓取视角搜索等方面开展研究工作。论文主要研究内容如下：

一、针对非结构化场景中6自由度抓取位姿检测中存在的位姿搜索空间大、部分观测等问题，提出一种基于投票机制的抓取位姿检测方法，通过投票机制缩小抓取位姿采样范围，并汇集点云不同区域的局部几何特征以实现抓取位姿的生成与评估。仿真实验与机器人实验结果表明基于投票机制的抓取位姿检测方法可以得到稳定且实际可执行的抓取位姿，能够较好地完成6自由度抓取任务。

二、针对仿真环境与真实机器人环境存在差异导致的仿真数据训练的模型在真实世界中性能下降的问题，提出一种基于领域自适应特征学习的抓取位姿检测方法，通过基于点云融合的对抗学习策略，促使特征提取器提取得到仿真数据和真实数据中具有一致性的特征，使得抓取位姿生成器能够统一地处理两个领域的特征信息。实验结果表明所提方法可以有效提升抓取位姿检测模型在目标域中的表现，同时可以适应不同的网络模型以及不同数据集之间的迁移，具有较好的可泛化性和实用性。

三、针对复杂堆叠场景中目标物体在单一观测视角下可见性较差导致无法有效抓取问题，提出一种基于抓取能力引导的机器人抓取视角搜索方法，综合目标物体的可见性以及可抓取能力设计奖励函数，通过强化学习的方式学习抓取视角搜索策略，引导机械臂主动地搜索对目标物体最优的抓取视角。机器人实验表明该视角搜索策略可以引导机械臂在少数几次视角优化后成功抓取初始观测中不可见的物体，具有较高的搜索效率与抓取成功率。

With the continuous expansion of robot applications from traditional industrial fields to service fields, new environments and tasks have put forward new requirements for the flexible grasping function of robotic arms.This thesis studies the methods of learning for grasping of daily objects in unstructured scenes, by conducting research on 6-DoF grasp pose detection, domain adaptation on feature learning and robot grasp viewpoint searching.The main contents of the thesis are summarized as follows:

(1) To address the issues of large search space and partial observation in 6-DoF grasp pose detection in unstructured scenes,a grasp pose detection based on voting mechanism is proposed in this thesis.Through voting mechanism, the sample space of grasp pose detection is narrowed,while local geometry feature from different regions of the point cloud is collected to achieve the generation and evaluation of grasp poses.Experiments conducted on both simulated and actual robots demonstrate that the performance of the grasp pose detection based on voting mechanism can obtain stable and practically executable grasp poses, and can effectively complete the taskof 6-DoF grasping.

(2) Aiming at the problem of performance decrease of models trained using simulated data when applied to real world,which is caused by the difference between the simulation and real robot environments,a grasp pose detection method based on domain adapted feature learning is proposed in this thesis.By an adversarial learning strategy based on point cloud fusion, the feature extractor is encouraged to extract the invariant features in the real and simulation data, allowing the grasp pose generator to process the feature information of the two domains evenly.Experimental results demonstrate that the proposed method can successfully enhance the network's performance in the target domain, and is also applicable to different network structures as well as the transference of data between diverse datasets, which demonstrated its good generalization and practicality.

(3) Aiming at the problem that the target object in a cluttered scene cannot be effectively grasped due to the partial visibility of the agent from a single viewpoint, grasp viewpoint searching approach based on grasping ability guidance is proposed in this thesis. Combining the visibility and graspability of the target object, the reward function is proposed, and the grasp view search policy is trained through reinforcement learning, to guide the active search of the optimal grasping viewpoint of the robotic arm.Robot experiments shows that this view search policy can guide the robotic arm to successfully grasp the target objects that are not visible in the initial observation after a few viewpoint optimizations, which shows the proposed method has high search efficiency and success rate of grasping.