Abstract Given the recent advances in computer vision, image processing and control systems, self-driving vehicles has been one of the most promising and challenging research topics nowadays. The design of vision-based robust controllers to keep an autonomous car in the center of the lane, despite uncertainties and disturbances, is still an ongoing challenge. This paper presents a hybrid control architecture that combines Deep Reinforcement Learning (DRL) and Robust Linear Quadratic Regulator (RLQR) for vision-based lateral control of an autonomous vehicle. Evolutionary estimation is used to model the vehicle uncertainties. For performance comparison, a DRL method and three other hybrid controllers are also evaluated. The inputs for each controller are real-time semantically segmented RGB camera images which serve as the basis to calculate continuous steering actions to keep the vehicle on the center of the lane with a constant velocity. Simulation results show that the proposed hybrid RLQR with evolutionary estimation of uncertainties architecture outperforms the other algorithms implemented. It presents lower tracking errors, smoother steering inputs, total collision avoidance and better generalization in new urban environments. Furthermore, it significantly decreases the required training time.

中文翻译：

---

基于深度强化学习的视觉鲁棒控制框架应用于自主地面车辆

摘要 鉴于计算机视觉、图像处理和控制系统的最新进展，自动驾驶汽车已成为当今最有前途和最具挑战性的研究课题之一。尽管存在不确定性和干扰，但设计基于视觉的稳健控制器以将自动驾驶汽车保持在车道中央仍然是一个持续的挑战。本文提出了一种混合控制架构，该架构结合了深度强化学习 (DRL) 和鲁棒线性二次调节器 (RLQR)，用于自动驾驶汽车的基于视觉的横向控制。进化估计用于对车辆的不确定性进行建模。为了进行性能比较，还评估了 DRL 方法和其他三个混合控制器。每个控制器的输入都是实时语义分割的 RGB 摄像头图像，这些图像作为计算连续转向动作的基础，以使车辆以恒定速度保持在车道中心。仿真结果表明，所提出的具有不确定性架构进化估计的混合 RLQR 优于其他实现的算法。它在新的城市环境中具有更低的跟踪误差、更平滑的转向输入、完全避免碰撞和更好的泛化。此外，它显着减少了所需的训练时间。它在新的城市环境中具有更低的跟踪误差、更平滑的转向输入、完全避免碰撞和更好的泛化。此外，它显着减少了所需的训练时间。它在新的城市环境中具有更低的跟踪误差、更平滑的转向输入、完全避免碰撞和更好的泛化。此外，它显着减少了所需的训练时间。