毕业生知识库

动力学与控制器学习是机器人领域主要的研究热点。现有方法大多是针对目标机器人采用小规模神经网络进行专有训练，通常具有泛化性差、精度不高等缺点。随着预训练技术的提出与发展，其在相关下游任务具有优异的表现，那么是否可以借助预训练技术以提高机器人建模和控制的泛化性、降低研究门槛？为此，本文将预训练引入机器人连续控制领域，研究串联机器人动力学和控制器预训练模型。

本文主要研究内容总结如下：

1.串联机器人动力学预训练。通过从模型维数、拓扑结构和动力学参数三个方面进行随机以产生大量异构串联机器人模型，再通过motor babbling随机若干运动轨迹，从而形成大规模的动力学轨迹数据集；利用Generative Pre-trained Transformer（GPT）结构学习大量机器人动力学以及三种目的不同的逆动力学模型。实验结果表明该方法在大规模轨迹测试集上具有较高的精度、较强的抗扰性以及泛化性，迁移到实际机器人时可以达到了与模型辨识相同的精度，并大幅降低了辨识过程的复杂度。

2.串联机器人控制器预训练。在上述大量异构串联机器人模型基础上，面向随机任务，优化全状态反馈控制器，并加入随机扰动以获得机器人的控制轨迹，从而构建大规模的运动控制轨迹数据集；利用GPT结构，以机器人随机动力学轨迹和运动控制轨迹为输入，学习大规模异构串联机器人的控制器预训练模型。实验结果表明该模型可以在仅给定动力学轨迹的先验下实现未辨识机器人和未知任务的闭环控制，具有较高的控制精度、鲁棒性和泛化性。

综上所述，本文将预训练引入机器人连续控制领域，构建了动力学和控制器的预训练模型，既展现了较高的精度、鲁棒性和泛化性，又能够降低相关领域研究门槛。

Dynamics and controller learning are one of the main research directions in the field of robotics. Most of the existing methods use small-scale neural networks to train specifically for a target robot, but they have the shortcomings of poor generalization and low accuracy. With the proposal and development of pre-training technology, it has excellent performance in related downstream tasks. A question is that can it be used to improve the generalization of robot modeling and control and lower the entry barrier? To this end, this thesis introduces pre-training into the field of continuous robot control and studies the pre-training models of dynamics and controller of serial articulated robots.

The major contents of this thesis are summarized as follows:

1. Dynamics pre-training of serial articulated robots. A large number of heterogeneous serial articulated robot models are generated by randomizing dynamics parameters, topology configurations, and

model dimensions, and then a large-scale dynamics trajectory dataset is formed by randomizing motion trajectories through motor babbling method. A structure modified from the generative pre-trained transformer(GPT) is applied to study the dynamics of massive robots and three inverse dynamics models with different purposes. The experimental results on large-scale trajectory test dataset show that our method has high accuracy, good performance in disturbance rejection and generalization. When transferred to an actual robot, it achieves the same accuracy as model identification , and greatly reduces the identification process complexity.

2. Controller pre-training of serial articulated robots. Given an above-mentioned random robot model, we optimize a full state feedback controller for a random task and add random disturbances to obtain a control trajectory. Iteratively repeating this process results in a large-scale motion control trajectory dataset. We also adopt the GPT structure to learn the controller pre-training model of large-scale heterogeneous series robots, taking a random dynamics trajectory and motion control trajectories as input. The experimental results show that the model can realize the closed-loop control of unidentified robot and unknown tasks with only a priori dynamics trajectory and has high control accuracy, robustness, and generalization.

In summary, this thesis introduces pre-training into the field of robot continuous control, and constructs pre-training models of dynamics and controller, which not only shows high accuracy, robustness and generalization, but also lowers the entry barrier of related area.