Robot automatic peg-in-hole assembly is a challenging task. Human perception-decision-operation skills play an irreplaceable role in precise assembly. In this paper, human assembly skills are modeled as a multi-level Markov decision process, and a multi-level imitation model is proposed to learn human assembly skills from demonstrations. Specifically, to learn human skills of identifying assembly phases based on perception signals, a perception-decision model based on parallel encoding Gaussian mixture model and decision correction module is proposed, and a new loss function combining reconstruction error and likelihood is designed. To accurately learn human operation skills from imprecise demonstrations, a decision-operation model combining mixture density networks and energy-based models is proposed. Based on the multi-level imitation model, a robot assembly controller is designed to drive robots to assemble like humans. Comparative experiments and peg-in-hole assembly experiments with four clearances indicate that the proposed method can more accurately learn human perception-decision-operation skills and achieve a better assembly effect than the advanced methods.

中文翻译：

---

像人一样组装：一种多层次的仿制模型，学习人类的感知-决策-操作技能，用于机器人自动组装任务


机器人自动钉入孔组装是一项具有挑战性的任务。人类的感知-决策-操作技能在精确装配中起着不可替代的作用。在本文中，将人体组装技能建模为多级马尔可夫决策过程，并提出了一种多级模仿模型，从演示中学习人体组装技能。具体来说，为了学习人类基于感知信号识别装配阶段的技能，提出了一种基于并行编码高斯混合模型和决策校正模块的感知决策模型，并设计了一种结合重建误差和似然的新型损失函数。为了从不精确的演示中准确学习人类操作技能，提出了一种混合密度网络和基于能量的模型相结合的决策-操作模型。基于多级仿制模型，设计了机器人装配控制器，驱动机器人像人类一样进行装配。对比实验和4个间隙的钉入孔组装实验表明，与先进方法相比，所提方法能够更准确地学习人类的感知-决策-操作技能，并取得更好的组装效果。