Abstract: To solve the problem of uncertain pose (position and pose) of peg and hole for improving the success rate,efficiency,and generalization ability of robot assembly operation,a variable admittance control strategy based on SARSA algorithm was proposed.The mechanism of peg-in-hole assembly was analyzed to guide the design of motion control strategy.Then,the different admittance parameters were analyzed by a series of simulations to obtain the controller parameters.In addition,an action evaluation method combining the displacement in each step and insertion depth was established to solve the problem of complex reward establishment in the assembly process.Meanwhile,qualification trace function is introduced to enhance learning efficiency of SARSA algorithm.Three sets of assembly experiments were implemented on the real robot.The improvement of success rate and efficiency could be verified by position and admittance control experiments.The performance of generalization ability was also validated through experiments in different initial pose.The results indicated that the proposed algorithm could solve the problem of the uncertain pose of peg and hole in the assembly task.

Key words: reinforcement learning, robot compliance control, peg-in-hole assembly, admittance controller

摘要： 为解决机器人装配任务中轴孔位姿不确定问题,提高装配操作的成功率、效率和泛化能力,提出一种基于SARSA算法的变参数导纳控制策略。首先,分析了轴孔装配机理,指导运动控制策略设计。其次,仿真分析了不同导纳参数的位置响应,获得控制器参数。此外,建立了以插孔深度和单次调整移动量为尺度的动作评估机制,解决装配过程奖励函数建模难题,并在动作价值更新过程引入资格迹函数提高算法学习效率。最后,在真实机器人上开展了3组实验,对比基于位置控制和导纳控制的实验结果验证了所提算法在装配成功率和效率上的提升,从不同初始位姿开始装配的实验结果验证了所提算法的泛化能力。结果表明,所提算法有望解决机器人装配任务中轴孔位姿不确定问题。

关键词: 强化学习, 机器人柔顺控制, 轴孔装配, 导纳控制器