Intelligent control for trustworthy autonomous vehicles

自动车辆控制对人工智能和控制理论提出了重大挑战。为了安全地控制自主车辆系统，控制系统需要了解在模型不确定性和环境扰动的影响下系统的动力学行为。这可以通过嵌入一种控制方案来实现，该控制方案包括：(1)基于模型的控制器的设计以确保系统的安全性；(2)在线学习技术的设计以估计未知系统的动态特性；(3)设计无模型学习技术(如强化学习)来估计未知环境。将基于模型的控制与在线系统动力学逼近和环境学习相结合，将使自动驾驶汽车在驾驶过程中对安全有高度的信心。因此，该项目的目标是设计一种控制器体系结构，该体系结构包括：(1)基于模型的控制器，利用安全关键控制概念；(2)未知系统动力学的在线学习；(3)无模型强化学习技术，用于估计外部环境，以确保学习和探索过程中的安全性。发展动态辨识技术，如高斯过程(GP)，对系统动态进行建模。开发基于(深度)强化学习的学习算法，以理解周围环境，即人类、障碍物。在Gazebo平台或MatLab/Simulink等计算机仿真平台上实现和验证所提出的算法。为了在无人机平台上实现所开发的算法，即HuskyA200 AGV机器人、PAWS机器人。

"Autonomous vehicle control presents a significant challenge for artificial intelligence and control theory. To safely control the autonomous vehicle systems, the control system needs to understand the behavior of system dynamics under the effects of model uncertainties and environmental disturbances. This can be achieved by embedding a control scheme that integrates: (1) the design of model-based controllers to ensure the safety of the system, (2) the design of online learning technique for estimating the unknown system dynamics, and (3) the design of model free-learning technique such as reinforcement learning to estimate the unknown environment. The integration between the model-based control with online system dynamics approximation and environment learning will enable the autonomous vehicle to highly have confidence on safety during driving. Therefore, the aim of the project is to design a controller architecture that combines: (1) a model-based controllers utilizing the safety critical control concepts, (2) an online learning of the unknown system dynamics, and (3) a model-free reinforcement learning technique, which estimates the external environment, to ensure safety during learning and exploration.The objectives of the project include:1. To design a model-based safety critical control that guarantees the autonomous vehicles will always operate within the safety zone.2. To develop dynamic identification technique, for example Gaussian Processes (GPs), to model the system dynamics.3. To develop learning algorithms based on (deep)-reinforcement learning to understand the surrounding environment, i.e, human, obstacles.4. To implement and validate the proposed algorithms in computer simulation platforms: Gazebo platform or Matlab/Simulink.5. To implement the developed algorithms in a UAV platform, i.e., HuskyA200 AGV robot, PAWs robot."