毕业生知识库

机器人是靠自身动力和控制能力来实现各种功能的一种机器。为了使机器人在真实的世界智能地与物体进行交互，感知，闭环的视觉控制以及灵活的操作都十分重要。其中，机器人灵活操作是机器人研究领域的重要研究问题之一，而稳定的物体抓取是机器人灵活操作的前提。基于此，本文通过力封闭分析方法以及深度强化学习方法对机器人稳定抓取进行了研究。

首先，由于深度确定性策略梯度（DDPG）算法可以处理连续控制问题，本文使用DDPG算法处理机器人抓取问题。针对DDPG算法样本使用效率低、学习速度慢、学习效果不理想等问题，融合多个DDPG算法的改进方法，如示教方法，数据优先重用机制等，以提高DDPG算法的性能及稳定性。

其次，分析了三指手稳定抓取力封闭方法。通过将DDPG算法与分析抓取稳定性的力封闭方法相结合，融入力封闭分析到DDPG算法的奖励机制中，实现了机器人对物体的稳定抓取。

最后，分别在仿真环境以及真实环境下对抓取方法进行了实验。在仿真实验中，针对本文提出的方法，从抓取成功率以及抓取稳定性两个重要方面进行了仿真实验。实验结果表明，机器人实现了接近90%的抓取成功率，并且抓取具有抗外力干扰的能力，实现了机器人抓取的稳定性。在实际环境的实验中，本文使用域随机化方法完成抓取策略从虚拟环境到真实环境的迁移，将在仿真中学习的策略直接应用在实体实验中，获得了较高的抓取成功率。

Robots are machines that can achieve various functions by their own power and control capabilities. Perception, closed-loop visual control and flexible operation are very important for robots to interact intelligently with objects in the real world. Among them, flexible operation of robots is one of the key issues in the field of robot research. Stable robotic grasping is the premise of robots'  flexible operation. Therefore, based on the deep reinforcement learning method and the force closure analysis method, this thesis aims to study stable robotic grasping.

First of all, as Deep Deterministic Policy Gradient (DDPG) algorithm can handle continuous control problem, this paper uses DDPG algorithm to deal with robotic grasping problem. Aiming at the problems of low sample efficiency, slow learning speed and unsatisfactory learning results of DDPG algorithm, this paper integrates multiple improved methods of DDPG, such as demonstrations and prioritized experience replay to improve the performance and stability of DDPG algorithm.

Secondly, the force closure grasping of three-fingered robot hand is analyzed. By combining DDPG algorithm with force closure method for analyzing grasping stability and integrating force closure analysis into the reward mechanism of DDPG algorithm, stable robotic grasping is achieved.

Finally, the proposed method was tested both in the simulation and real environment. In the simulation experiment, the proposed method was tested in two important aspects of robotic grasping: the grasp success rate and the grasp stability. The experimental results show that the proposed method has achieved nearly 90% grasp success rate and the grasps have the ability to resist external force to achieve the stability of robotic grasping. In real world experiments, this thesis uses domain randomisation method to fulfill the transfer of the learned policy from simulation to the real environment. The policy learned in the simulation was directly applied to the real environment and got good grasp success rate.