基于深度强化学习的双向装配序列规划

Assembly sequence planning based on deep reinforcement learning

为了解决复杂装配模型的序列规划问题,并使算法对任意初始状态具有较高的适应性,本文提出了一种包含正向装配以及逆向拆解的一体化双向装配序列规划方法BASPW-DQN.针对复杂装配模型,首先进行了一体化装配序列规划的问题描述与形式化表示;在此基础上,引入了课程学习及迁移学习方法,对包含前向装配和逆向错误零件拆卸两部分过程的双向装配序列规划方法进行研究.在所搭建的ROS-Gazebo与TensorFlow相结合的仿真平台上进行了验证,测试结果证明此双向网络对于任意初始状态(包括零装配、部分装配、误装配等初始状态)的装配任务均可以在较少步数内完成,验证了所提方法对于解决装配序列规划问题的有效性与适应性.

In order to solve the sequence planning problem of complex assembly models and improve the flexibility of the algorithm to any initial state,this paper proposes an integrated bi-directional assembly sequence planning method BASPW-DQN.Aiming at the complex assembly model,a bi-directional assembly sequence planning method including forward assembly and wrong part disassembly process is proposed,on this basis,curriculum learning and transfer learning methods are introduced to improve the training efficiency and assembly capabilities of the integrated assembly sequence planning method.And a training platform is developed,which combines the physical simulator Gazebo and deep network framework TensorFlow.The test results show that the bi-directional network can complete the assembly tasks of general assembly in any initial state(such as none-assembly,partial assembly and misassembly)in a few steps demonstrating the effectiveness and flexibility of the proposed method.