Abstract: To solve the problems such as low training efficiency and poor adaptability due to the heavy reliance on manual teaching data,a reinforcement learning strategy considering variable rotation parameters was proposed for robotic multiple peg-in-hole assembly.An attitude adjustment model with variable rotation parameters was presented to collect the corresponding relationship data between the contact mechanics information of multiple peg-holes and the attitude adjustment action,from which the data could be used as the pre-training learning data of assembly skills.Furthermore,an improved Deep Deterministic Policy Gradient(DDPG)reinforcement learning algorithm was proposed,which used multi-factor sparse reward function to perform appropriate reward evaluation for assembly actions to improve learning efficiency and success rate.Finally,a case study of multiple peg-in-hole electronic component assembly was carried out on the simulation and experimental platform.The results showed that the proposed strategy had good scene adaptability,and could effectively improve the learning efficiency and success rate of assembly compared with the classical reinforcement learning method,while significantly reducing the contact force/torque of assembly.

Key words: collaborative robot, multiple peg-in-hole assembly, attitude adjustment model, improved deep deterministic policy gradient algorithm

摘要： 针对目前机器人多轴孔装配学习策略严重依赖人工示教数据,导致训练效率低和场景适应性差等问题,提出一种考虑可变旋转参数的机器人多轴孔装配强化学习策略。首先,提出一种可变旋转参数的姿态调整模型,据此采集多轴孔接触力学信息与姿态调整动作的对应关系数据,以此作为装配技能的预训练学习数据。进而,提出一种改进深度确定性策略梯度(DDPG)强化学习算法,通过多因素稀疏奖励函数对装配动作进行合适的奖励评价以提高学习效率和成功率。最后,在仿真和实验平台上进行了多轴孔电子元器件装配的案例研究,结果表明,所提方法具有良好的场景适应性,相对经典强化学习方法能有效提高装配的学习效率和成功率,同时明显减小了装配接触力/力矩。

关键词: 协作机器人, 多轴孔装配, 姿态调整模型, 改进深度确定性策略梯度算法