近年来，无人驾驶技术在汽车领域迅猛发展，自行车作为和汽车一样被人们广泛运用的交通工具，自行车的无人驾驶也备受关注。自行车是一个不稳定系统，其平衡控制是无人驾驶自行车的基础，也是其面临的最大挑战之一。自行车平衡控制不仅需要实现车身的平衡，而且需要实现灵活的转向和变速，以应对自行车在实际路面遇到的突发状况。 本文采用车把转向控制方法实现前轮转向、后轮驱动结构自行车的平衡控制，该方法利用了自行车自身的转向动力学特性，具有能量消耗低和反应速度快的优点。本研究中搭建了自行车的样机，根据样机建立了自行车转向动力学的非线性模型，分析了动力学模型的开环特性并定义了自行车平衡点的概念，为车把转向控制器的设计提供理论依据。为了降低控制器的设计难度，先忽略车速变化对模型的非线性影响，设计出恒速车把转向控制器；在恒速车把转向控制器基础上，进一步基于速度分段方法设计了变速车把转向控制器，实现了无人驾驶自行车的灵活变速；最后基于前馈神经网络设计出变速车把转向控制器，通过实验验证其可行性。 论文在展望部分提出了电机控制方法以及速度测量等方面的改进方案，同时在神经网络的控制器基础上利用机器学习实现自行车平衡策略的自主学习，期望进一步提升无人驾驶自行车的性能。

Recently, the technology of autonomous driving has been developing rapidly. Similar to cars, which are widely used in transportation, bicycles’ autonomous driving is also under the spotlight. The bicycle is an unstable system, and the balance control of bicycles is the basis of and one of the hardest challenges faced by autonomous bicycles. The balance control is not only required to maintain the balance of the bicycle’s body, but also to achieve flexible turning and velocity shifting, in order to deal with unexpected situations on the actual road. In this thesis, the steering control method is applied to achieve the balance control of the bicycle, assuming a front-wheel steering and rear-wheel driving bicycle structure. The steering control method takes advantage of the bicycle steering dynamics, resulting in low energy consumption and fast response. This research introduces the bicycle prototype and the nonlinear dynamics model of the bicycle and the open-loop characteristics of the dynamics model are analyzed. To provide a theoretical basis for the design of the steering controller, the concept of equilibrium points is defined according to the characteristics of the bicycle system. In order to reduce the difficulty of the controller design, a constant-velocity steering controller is designed at first, ignoring the non-linear change of the dynamics model brought by the change of bicycle velocity. Based on the proposed constant-velocity steering controller, the varying-velocity steering controller is designed with the method of velocity segmentation; Finally, the implementation of a varying-velocity steering controller with the use of feedforward neural networks is proposed, and the experiments on the bicycle prototype prove the feasibility of the controller. In addition, in order to improve the performance of the bicycle prototype, some modifications are proposed as future research extensions. These include improvements on the control method of the steering motor, on the velocity measurement, and on other aspects, as well as applying the machine learning methodology to realize the self-learning of bicycles' balance control strategies based on feedforward neural networks.