中国科学院自动化研究所机构知识库

精密装配可以实现高精度的操作，在军事、航天等国家重要战略性领域内有不可替代的作用。从某种意义上讲，机器人精密装配技术决定了部分高端国防装备的性能。精密装配目前还存在抗扰能力差、自动化程度不高等诸多问题。引入多维被动柔顺装置可以有效提高装配柔顺性、降低接触力以及保护微器件；智能化装配可以提升效率、减少人为参与。因此，如何赋予精密装配机器人基于多维被动柔顺装置的操作学习能力，已成为精密操作中提高机器人智能性的关键科学问题之一。本文着眼于精密装配的抗扰控制与技能学习，引入多维被动柔顺装置，探讨面向姿态自适应的抗扰控制策略，并以深度强化学习为主要手段，提出被动柔顺精密装配的技能学习与迁移方法。本文主要研究工作如下：

1. 引入多维被动柔顺装置的精密装配抗扰控制。构建多维被动柔顺装置的形变-接触力模型，提出常规条件与姿态受扰情况下的控制策略；充分利用轴向柔顺，提出再调整策略，以应对卡阻等异常现象；再综合进给预测、接触力以及估计误差等因素，提出面向多种不确定性的高效运动规划，最终实现了高效稳定的被动柔顺精密装配抗扰控制。该策略能应对的姿态扰动可达7度，可降低约90%的径向作用力，大幅提高装配成功率。

2. 面向被动柔顺精密装配的技能学习。利用编码解码网络学习多维被动柔顺形变-接触力模型的分布参数；将该柔顺模型与强化学习TD3算法相结合，以模型的统计特征作为动作调节参数，提出适用于被动柔顺精密装配的技能学习算法，从而实现高效稳定的技能学习。相比前述的被动柔顺控制策略，该方法在真实环境学习5小时的技能可进一步降低62%的径向力，提高操作性能。

3. 面向被动柔顺精密装配的技能仿真-现实迁移。在仿真中学习技能、再迁移到现实中，可以有效降低深度强化学习在真实场景训练的时间成本与硬件成本。为此，首先使用环境集合构建仿真环境，以应对多维被动柔顺模型的不确定性，再结合元强化学习与被动柔顺模型网络，实现被动柔顺精密装配的技能迁移。在线训练100分钟即可完成技能迁移，装配效果与前述技能相当，但训练用时减少了66%，并大幅降低算法训练成本。

Precision assembly can achieve high-precision operations, and plays an irreplaceable role in national strategic areas such as military and aerospace. In a sense, robot precision assembly technology determines the performance of some high-end national defense equipment. At present, precision assembly still has many problems such as poor disturbance rejection ability and low degree of automation. The introduction of multi-dimensional passive compliant devices can effectively improve assembly compliance, reduce contact force and protect micro devices; intelligent assembly can improve efficiency and reduce human involvement. Therefore, how to enable precision assembly robots to learn operations based on multi-dimensional passive compliant devices has become one of the key scientific issues for improving robot intelligence in precision operations. This thesis focuses on the disturbance rejection control and skill learning of precision assembly, and by introducing multi-dimensional passive compliant devices, discusses disturbance rejection control strategies for posture adaptation. It proposes skill learning and transfer methods for passive compliant precision assembly on the basis of deep reinforcement learning. The main research work of this thesis is as follows:

1. Disturbance rejection control of precision assembly with multi-dimensional passive compliant device. This thesis proposes control strategies under normal conditions and posture disturbances by constructing the deformation-contact force model of the multi-dimensional passive compliant device. To make full use of the axial compliance, it proposes readjustment strategies to deal with abnormal phenomenas such as jamming. In this paper, insertion forecasting, contact force and estimation error are integrated, and an efficient motion planning for multiple uncertainties is proposed, and finally an efficient and stable passive compliant precision assembly disturbance rejection control is realized. This strategy can deal with attitude disturbances of up to 7 degrees, which can reduce about 90% of the radial force and greatly increase the success rate of assembly.

2. To learn the skills of passive compliant precision assembly. This thesis uses the encoder-decoder network to learn the distribution parameters of the multi-dimensional passive compliant deformation-contact force model. By combining the compliant model with the reinforcement learning TD3 algorithm and using the statistical characteristics of the model as the action adjustment parameters, it proposes a skill learning algorithm suitable for passive compliant precision assembly to achieve efficient and stable skill learning. Compared with the aforementioned passive compliant control strategy, this method can further reduce the radial force by 62% and improve the operation performance by learning the skills for 5 hours in the real environment.

3. Skill transfer from simulation to reality for passive compliant precision assembly. Learning skills in simulation and then transferring to reality can effectively reduce the time cost and hardware cost of deep reinforcement learning training in real scenarios. To this end, the environment set is first used to build a simulation environment to deal with the uncertainty of the multi-dimensional passive compliant model, and then combined with meta-reinforcement learning and passive compliant model network to realize the skill transfer of passive compliant precision assembly. Experiments show that the skill transfer can be accomplished whih only 100 minutes of online training. The assembly effect is equivalent to the aforementioned skills, but the training time is reduced by 66%, which greatly reduces the cost of algorithm training.