当前的底盘控制系统主要通过分配各车轮制动力与驱动力防止车轮抱死或打滑，以维持车身稳定性。但在专业驾驶比赛中，驾驶员会通过有意识地控制车轮抱死或打滑完成漂移等极限驾驶策略，以减小圈时或躲避障碍物。当前底盘控制系统受限于传统车辆感知范围的局限性，控制算法仅基于自车状态信息，无法利用周围环境的道路、障碍物等信息。因此，通过研究极限工况下的车辆动力学特性和控制策略，可以指导下一代自动驾驶车辆底盘控制系统的设计，扩展其控制边界。本文针对不同的极限驾驶工况，提出了稳态漂移工况的平衡态稳定控制方法、极限和非极限混合工况的轨迹规划和运动控制方法以及基于人工智能强化学习的车辆极限工况控制方法。首先，提出了考虑有限控制能力和模型不确定性的稳态漂移控制方法。建立稳态漂移工况下的车辆动力学模型和考虑附着圆限制的非线性轮胎模型，分析稳态漂移工况的动力学产生机理。基于线性二次型算法设计前馈与反馈相结合的稳态漂移控制器，针对稳态漂移工况下轮胎附着力接近饱和导致的有限控制能力和路面附着系数变化导致的模型有界不确定性，设计了系统的最大吸引域和稳态漂移状态不变集，保证系统状态处于吸引域内时可以收敛于不变集。其次，提出了适用于极限和非极限混合工况的轨迹规划和运动控制方法。轨迹规划算法将可行驶区域分为不同类型，在极限工况下采用基于规则和随机采样算法规划期望轨迹，在非极限工况采用结合车辆动力学模型的快速扩展随机树算法规划期望轨迹。运动控制算法基于单轨车辆模型和线性轮胎模型，采用开闭环相结合的控制策略，利用模型预测算法设计控制模式的切换策略，实现漂移参考轨迹的跟踪。然后，提出了极限工况下结合先验知识的强化学习算法。基于Actor-Critic强化学习框架，设计了反映系统控制律的行为网络和评估控制律优劣的评价网络，通过策略梯度算法和时序差分算法更新行为网络和评价网络。强化学习算法分别以人工演示和最优控制算法为先验知识，有助于行为网络和评价网络收敛至全局最优，经过训练的神经网络达到了专业驾驶员在极限工况下的操控能力。最后，搭建了1:10比例的实车平台，完成了稳态漂移和瞬态漂移实车实验。实验结果表明了本文提出的极限工况下轨迹规划和运动控制算法的可行性。
Current chassis control systems try to prevent tires from locking or slipping to maintain vehicle stability. In racing instead, expert drivers intentionally perform drifting maneuvers by making tire forces saturating to reduce lap time or avoid obstacles. Current chassis control systems are subject to the limited sensoring ability of traditional vehicles and could only make use of ego vehicle information. By studying vehicle dynamics and control maneuvers in extreme conditions, we can help design advanced chassis control systems with an extended operation envelope. Based on the vehicle dynamics in different extreme scenarios, we propose linear quadratic equilibrium controller for sustained drift, trajectory planning and motion control algorithms to plan and track a reference drift trajectory for transient drift, and vehicle extreme handling maneuvers based on reinforcement learning. Firstly, we design an equilibrium controller with considering limited control authority and model uncertainty. A single-track bicycle model and a nonlinear tire model with considering friction circle limit are proposed. We analyze the vehicle dynamics under sustained drift and design a linear quadratic sustained drift controller with a mixed feedforward and feedback control structure. To deal with the limited control authority caused by tire nearly saturating and model uncertainty caused by friction coefficient’s variation, we design the maximum attraction region and minimum invariant set. Secondly, a complete trajectory planning and motion control framework is proposed, which is suitable for combined extreme and non-extreme scenario. The trajectory planner divides the path horizon into different types of regions. The ruled-based sampling method is applied to find a path in extreme region, and the Rapidly-exploring Random Tree with single-track model is applied to find a path in non-extreme region. A mixed open-loop and closed-loop control technique based on the bicycle model with linear tire model is applied to track the drift trajectory. The switch policy is based on the model predict method. Next, a reinforcement learning algorithm with prior knowledge is proposed to achieve transient drift under extreme conditions. Based on Actor-Critic reinforcement learning architecture, an actor neutral network is designed to reflect the control policy and a critic neutral network is designed to evaluate the control policy. The actor network is updated through policy gradient method and the critic network is updated through Temporal-Difference method. By learning from the demonstrations and optimal control policy, the prior knowledge is applied to help actor network and critic network converge to the global optimum. The resultant control strategy is consistent with the expert drivers’ behaviors performed in racing.Finally, a 1:10 real vehicle platform is constructed to verify the proposed methods. The experimental results show the proposed trajectory planning and motion control algorithm are applicable in practice.
