毕业生知识库

随着人工智能技术在感知、决策等领域取得关键性突破，应用了各类先进智能算法的机器人系统，有望实现高自主性、高柔性的机器人作业，展现出更高的智能化水平。本文面向机器人作业场景，针对机器人操作中的抓取任务学习、轴孔装配技能学习、线缆布线技能学习及复杂长序任务等方面展开研究工作，论文的主要内容如下：

一、针对基于视觉感知的机器人抓取任务学习问题，提出了一种视觉示教编程方法。首先根据图像特性设计实虚图像变换函数预处理真实图像，通过减少数据分布差异提升了基于仿真数据训练的模型在真实图像上的检测性能。然后，设计了一种技能识别算法以从人类示教视频中生成抓取任务的机器人作业程序。仿真与真实环境下的机器人抓取作业实验验证了所提方法的有效性。

二、针对机器人轴孔装配技能学习问题，提出了一种基于元强化学习的机器人技能学习方法。通过设计任务条件残差策略划分了任务共性知识，提高了元强化学习算法在元训练期间的样本效率。然后，利用任务提取网络从人类示教推理出任务特定知识，并与任务条件残差策略中的任务共性知识融合提高了元强化学习算法在元测试阶段的样本效率。通过多种不同类型的机器人轴孔装配实验验证了所提方法的有效性。

三、针对机器人线缆布线技能学习问题，提出了基于自监督的机器人零样本模仿学习方法。在自监督策略学习阶段，设计了目标一致性损失函数以避免训练多步目标条件策略时可行预测动作被错误惩罚，从而提高了训练策略的作业性能。在机器人模仿作业阶段，提出了一种关键帧提取方法，通过评估策略在示教路径上的作业性能构建了由所有可行路径组成的有向图，并利用最短路算法提取了策略在该任务上的最优关键帧。通过机器人线缆布线、线缆构形等作业实验验证了所提方法的有效性。

四、针对机器人复杂长序作业问题，设计并实现了基于自主与模仿融合的机器人作业系统。首先，使用前面章节提出的自主学习方法分别获取机器人抓取、轴孔装配以及线缆布线等任务的感知模型或控制策略。然后，在已习得模型和策略基础上定义可复用的机器人技能，这些技能可利用示教和任务知识快速迁移到特定场景，并且通过技能组合可以完成复杂长序作业任务。在照明电路接线场景下机器人作业实验验证了所设计系统的有效性。

As artificial intelligence technology has made breakthroughs in perception and decision-making, the robot system with advanced intelligent algorithms is expected to achieve high autonomy and flexibility in robotic manipulations and exhibit a higher level of intelligence. This thesis aims at robot manipulations and conducts research on pick-and-place tasks, peg-in-hole assembly, cable routing, and complex long-horizon tasks. The main contents of this thesis are as follows:

Firstly, programming by visual demonstrations is proposed to tackle the problems of pick-and-place task learning. A sim-to-real image mapping function is designed to preprocess the image to reduce the difference in data distribution, which improves the real-world detection performance of the model trained by the simulated data. Then, a skill recognition algorithm is designed to generate a robot program from a visual demonstration. The experiment in simulated and real-world environments verifies the effectiveness of the proposed method.

Secondly, a meta-reinforcement learning method is proposed for the problem of robot learning peg-in-hole assembly. The task-shared knowledge is divided by designing a task-conditional residual policy, improving sample efficiency during meta-training. Then, task-specific knowledge is inferred from human demonstrations by the task extraction network. And the sample efficiency during meta-testing has been improved by fusing task-shared and task-specific knowledge. Several different robotic peg-in-hole assembly experiments verify the proposed method's effectiveness.

Thirdly, a robot zero-shot imitation learning method based on self-supervised learning is proposed to tackle the problem of learning cable routing. In the self-supervised policy learning phase, goal consistency loss is designed to avoid penalizing feasible predicted actions when training multi-step goal-conditioned policy, improving the task performance of the learned policy. An estimation-based keyframe extraction method is proposed in the robot imitation phase to improve robot manipulation efficiency. A directed graph of all feasible paths is constructed by predicting the trained policy's performance on any transition pair in the demonstration. Then, the shortest path algorithm extracts the task's optimal keyframes for the learned policy. The effectiveness of the proposed method is verified by robotic manipulation experiments such as cable routing and configuration.

Finally, aiming at the problem of complex robotic long-horizon manipulations, a robot manipulation system based on the fusion of autonomy and imitation is designed and implemented. First, the methods proposed in previous chapters are used to learn the perception models or control policies for robotic grasping, peg-in-hole assembly, and cable routing. Then, three reusable robot skills are defined based on the learned models and policies. These skills can be quickly transferred to specific scenarios by demonstration or task knowledge. And the robot can complete complex long-horizon tasks by combining designed skills. The robot manipulation experiment in the lighting circuit wiring scene verifies the effectiveness of the designed system.