伴随着“十四五”规划和“中国制造2025”的预期蓝图不断推进，我国产业智能化水平也在不断提升，智能机器人逐步迈入日常生活并协同人类完成一些操作任务，但大多数机器人高刚度和强驱动的特点使得人们不得不重视协同操作过程中的安全问题。传统的工业机器人往往处在相对整洁的结构化车间，而人机协同操作任务大多处于动态非结构化环境，环境中物体种类繁多且随机分布，甚至与人类的活动空间重合。此时，机器人很难像在工业车间那样精确地获取周围所有物体的位姿信息与运动状态来保障操作安全。为此，本文针对机器人在人机协同操作中的安全问题展开研究，具体研究内容如下： 首先，为了给机器人提供良好的接触反馈以应对非结构动态环境中障碍物引发的意外接触，本文设计一款具备视触觉反馈的多维弹性夹爪。该夹爪借助一个片状的感知弹性体作为与被操作物体接触的媒介并通过一个内置相机观察弹性体形变作为视触觉反馈信息。同时，弹性体也可以作为接触缓冲器，为机器人的柔顺操作奠定基础。 其次，为了保障在复杂的环境中机器人仍可以稳定抓取目标物体，本文提出了一种逐点注意力投票网络PAV-Net。该方法将注意机制引入到关键点的投票过程中从而高效整合数量不等的可见点特征，降低复杂环境带来的干扰。PAV-Net在LineMOD和YCB-Video两个公开数据集中展现了优异的实时位姿估计性能。 随后，基于多维弹性夹爪的接触模型，本文提出一种柔顺接触控制方法。该方法通过对感知弹性体形变的分析建立一个针对弹性体形变的力-位移模型并以此估计夹爪与被操作物体间的触觉力和力矩。此外，在面对由未知障碍物引起的意外接触时，该方法可以主动地调整弹性体的形变程度以保证接触的柔顺性与安全性，且在面对障碍物阻碍操作时，其也可以让被操作物体沿着障碍物表面完成规避。 最后，一系列的仿真和实物实验证明了本文设计的夹爪与提出的安全接触控制在面对由未知障碍物与被操作物体发生意外接触时具备良好的柔顺性和鲁棒性，即便是主动靠近的动态障碍物也可以实现稳定接触并规避。此外本文还模拟了人机协同操作时机器人误触人类的实验场景。实验结果表明手持刀具的机器人在完成操作任务后即便误触到人也不会造成误伤，证明了本文方法的安全性。

With the continuous progress of the "14th Five-Year Plan" and the expected blueprint of "Made in China 2025", the level of intelligence in Chinese industry is also increasing, and intelligent robots are beginning to enter daily life and help people complete some manipulation tasks. However, the high stiffness and strong actuation properties of most robots make people pay attention to safe interaction during human-robot cooperative manipulations. Traditional industrial robots are often located in organized workshops, while most of human-robot cooperative manipulation tasks are in the dynamic unstructured environments where objects are diverse and randomly distributed, and even overlap with the human activity space. At this time, it is difficult for the robot to accurately obtain the positional information and motion status of all the surrounding objects as in the industrial workshop to ensure the safety of manipulations. Therefore, this thesis focuses on the safe contact of robot in human-robot cooperative manipulations, and the specific research contributions are as follows: First, to provide excellent contact feedback to the robot to cope with unexpected contact caused by obstacles in unstructured dynamic environments, this thesis designs a multi-dimensional elastic gripper with visual-tactile feedback. The gripper uses a sheet-shaped perceptual elastomer as the contact medium with the manipulated object and observes the deformation of the elastomer through a built-in camera as the visual-tactile feedback information. Meanwhile, the elastomer can also be used as a contact cushion, laying the foundation for compliant manipulation. Second, this thesis proposes a point-wise attention voting network, PAV-Net, to ensure that the robot can stably grasp the target object even in a complex environment. The method introduces the attention mechanism into the voting process of key points to realize the efficient integration of a varying number of visible point features and reduce the interference caused by the complex environment. PAV-Net demonstrates excellent real-time pose estimation performance in two publicly available datasets, LineMOD and YCB-Video. Third, based on the contact model of the multidimensional elastic gripper, this thesis proposes a compliant contact control method. By analyzing the deformation of the elastomer, the method establishes a force-displacement model of the elastomer deformation, which can estimate the tactile force and torque between the gripper and the manipulated object. In addition, the method can actively adjust the elastomer deformation to ensure the compliant and safe interaction in the face of unexpected contact caused by unknown obstacles, and it can also realize automatic avoidance along the surface of the obstacle in the face of serious obstacles that hinder the manipulation. Last, a series of simulations and physical experiments demonstrate that the designed gripper and the proposed safe contact control in this thesis has excellent flexibility and robustness in the face of unknown obstacles making unexpected contact with the manipulated object. Even the dynamic obstacles actively approaching the robot can stably contact and avoid them. Moreover, this thesis also simulates an scenario in which a robot mistakenly contacts a person during a human-robot cooperative manipulations. The experiment results show that the robot holding a knife will not cause dangerous injury even if it mistakenly contacts a person after completing the manipulation task, which proves the safety of the method in this thesis.