A reinforcement learning (RL) based method that enables the robot to accomplish the assembly-type task with safety regulations is proposed. The overall strategy consists of grasping and assembly, and this paper mainly considers the assembly strategy. Force feedback is used instead of visual feedback to perceive the shape and direction of the hole in this paper. Furthermore, since the emergency stop is triggered when the force output is too large, a force-based dynamic safety lock (DSL) is proposed to limit the pressing force of the robot. Finally, we train and test the robot model with a simulator and build ablation experiments to illustrate the effectiveness of our method. The models are independently tested 500 times in the simulator, and we get an 88.57% success rate with a 4mm gap. These models are transferred to the real world and deployed on a real robot. We conducted independent tests and obtained a 79.63% success rate with a 4mm gap. Simulation environments: https://github.com/0707yiliu/peg-in-hole-with-RL.

翻译：提出了一种基于强化学习的机器人装配方法，该方法能够在安全约束下完成装配任务。整体策略包含抓取与装配两个阶段，本文主要聚焦于装配策略的设计。为感知孔洞的形状与方位，本文采用力反馈替代视觉反馈。此外，由于过大输出力会触发急停机制，提出了一种基于力的动态安全锁（DSL）以限制机器人的压入力。最后，通过仿真器对机器人模型进行训练与测试，并构建消融实验以验证本文方法的有效性。模型在仿真器中独立测试500次，在4毫米间隙条件下成功率达88.57%。将该模型迁移至真实场景并部署于实体机器人，经独立测试，在4毫米间隙条件下成功率为79.63%。仿真环境参见：https://github.com/0707yiliu/peg-in-hole-with-RL。