针对混杂物体散乱堆叠下的机器人抓取场景，使用固定视角相机的视觉抓取存在成功率低的问题，提出一种基于深度强化学习框架的眼-手随动相机视角选择策略，令机器人能够自主地学习如何选择合适的末端相机位姿，以提高机器人视觉抓取的准确率和速度。首先，面向机器人主动视觉抓取任务建立马尔科夫决策过程模型，将视角选择问题转化为对视角价值函数的求解问题。使用编码解码器结构的反卷积网络近似视角动作价值函数，并基于深度Q网络框架进行强化学习训练。然后，针对训练过程中存在的稀疏奖励问题，提出一种新的视角经验增强算法，分别对抓取成功和抓取失败的过程设计不同的增强方式，将奖励区域从单一点拓展到圆形区域，提高了视角动作价值函数近似网络的收敛速度。先期实验部署在仿真平台中，通过搭建机器人模型及仿真抓取环境实施离线强化学习训练。过程中，使用提出的视角经验增强算法可以有效提高样本利用率，加快训练的收敛速度。基于所提出的视角经验增强算法，视角动作价值函数近似网络在2 h以内可达到收敛。为验证所提视角选择策略的实际应用效果，将视角经验增强算法实施在真实场景下的机器人主动视觉抓取实验中。实验结果表明，采用该策略进行的视角优化有效提高了机器人的抓取准确率和抓取速度。相较其他方法，所提出的视角选择策略在实际机器人抓取中只需进行一次视角选择即可获得抓取成功率高的区域，进一步提高了最佳视角选择的处理效率。相对于单视角方法，混杂场景的抓取成功率提升22.8%，每小时平均抓取个数达到294个，具备了进入工业应用的可行性。

To solve the problem of the low success rate of robot vision grasping using fixed environment camera in the scene of cluttered and stacked objects, an eye-hand follow-up camera viewpoint selection policy based on deep reinforcement learning is proposed to improve the accuracy and speed of vision-based grasping. Firstly, a Markov decision process model is constructed for robot active vision-based grasping task, then the problem of viewpoint selection is transformed into a problem of solving the viewpoint value function. A deconvolution network with encoder-decoder structure is used to approximate the viewpoint action value function, and the reinforcement learning is carried out based on the deep Q-network framework. Then, to resolve the problem of sparse reward existing in reinforcement learning, a novel viewpoint experience enhancement algorithm is proposed. The different enhancement methods between the successful and failed grasping process are designed respectively. And the reward region can be expanded from a single point to a circular region for improving the convergence speed of the approximation network. The preliminary experiment is deployed on the simulation platform, and the robot model and the grasping environment are simultaneously built in the simulation platform to implement the offline reinforcement learning. In the process, the proposed viewpoint experience enhancement algorithm can effectively improve the sample utilization rate and speed up the convergence of training. Based on the proposed viewpoint experience enhancement algorithm, the viewpoint action value function approximation network can converge within 2 h. To obtain the results from the verification with application, the proposed viewpoint selection policy is applied to the real-world scenes with robot for grasping experiments. The result shows that the viewpoint optimization based on this policy can effectively promote the accuracy and speed of robot grasping. Compared with the general grasping methods, the proposed viewpoint selection policy needs only one viewpoint selection in real-world robot grasping to find the focus region with high grasping success rate. And the method can also promote the processing efficiency of the best viewpoint selection. The grasping success rate in cluttered scenes is increased by 22.8% against the single-view method, and the mean picks per hour can reach 294 units. As whole, it shows that the proposed policy has the capacity of industrial application.