Learning of safety critical model predictive controllers for autonomous systems

诸如移动的机器人和自主车辆的现代自主系统在很大程度上依赖于用于运动控制的反馈控制器，特别是用于路径跟随和障碍物回避，其中它们被用来跟随由分层控制方案中的更高级别的运动规划器设置的轨迹。模型预测控制（MPC）是用于避障的流行控制器选择，因为它允许指定约束以确保移动的机器人或自主车辆不与障碍物碰撞。MPC的行为在多年的理论发展和工业实践中得到了很好的理解，提供了强有力的安全保证，但需要大量的时间和专业知识来实现它，特别是在自动驾驶汽车等安全关键应用中。近年来，深度强化学习（RL）的研究为自动寻找非线性反馈控制器提供了新的方法，以解决具有挑战性的控制问题。但与MPC不同的是，现有的RL方法通常无法保证稳定性或约束满足，并且对于安全关键应用，很难验证它们的行为。为了将MPC的可预测性和安全性保证与现代RL方法的功能和便利性相结合，该项目将开发在Actor-Critic RL框架中自动学习MPC控制器的方法，考虑自动驾驶汽车的运动控制和避障问题。这将是最近的数学结果的直接应用，表明凸优化（如MPC）可以用作PyTorch等RL框架中的可训练层，允许它们被学习。目标是实现路径跟踪型MPC的快速设计和原型化，而无需底层MPC算法的专家知识，从而减少开发时间和成本，并提高未来移动的机器人和自动驾驶汽车的安全性和可靠性。为了确保新算法的实际应用，摩托车路径跟踪和稳定性辅助的示例应用将用于指导其开发。例如通过转向致动来稳定向前运动中的两轮车辆以遵循预定义路径的问题是具有挑战性的，并且在摩托车和踏板车的主动安全系统的新兴领域中具有重要应用。为了产生长期影响并鼓励自治系统研究人员采用新方法，开发的新方法将包含在Github上发布的开源软件库中。

Modern autonomous systems such as mobile robots and autonomous vehicles rely heavily on feedback controllers for motion control, particularly for path-following and obstacle avoidance, where they are employed to follow a trajectory set by a higher-level motion planner in a hierarchical control scheme. Model Predictive Control (MPC) is a popular controller choice for obstacle avoidance, as it allows constraints to be specified to ensure that the mobile robot or autonomous vehicle does not collide with obstacles. The behaviour of MPC is well understood from years of theoretical development and industrial practice, providing strong safety assurances, but considerable time and expert knowledge is required to implement it, especially in safety-critical applications such as autonomous vehicles.In recent years, research on deep Reinforcement Learning (RL) has provided new methods to automatically find nonlinear feedback controllers for challenging control problems. But unlike MPC, existing RL methods typically have no guarantees of stability or of constraint satisfaction, and for safety-critical applications it is difficult to verify their behaviour. To combine the predictability and safety guarantees of MPC with the power and convenience of modern RL methods, this project will develop methods to automatically learn MPC controllers in actor-critic RL frameworks, considering motion control and obstacle avoidance problems for autonomous vehicles. This will be a direct application of recent mathematical results showing that convex optimisations, such as MPC, can be employed as a trainable layer in RL frameworks such as PyTorch, allowing them to be learned. The goal is to enable rapid design and prototyping of path-following type MPC without requiring expert-knowledge of the underlying MPC algorithm, therefore reducing development time and cost and improving safety and reliability of future mobile robots and autonomous vehicles.To ensure the new algorithms are practically applicable, an example application of motorcycle path-following and stability assistance will be used to guide their development. The problem of stabilising a two-wheeled vehicle in forward motion to follow a predefined path, for example via steering actuation, is challenging and has important applications in the emerging area of active safety systems for motorcycles and scooters. For long term impact and to encourage adoption of the new methods by autonomous systems researchers, the new methods developed will be included in an open-source software library published on Github.