CAREER: Real-time Convex Optimization for High-Performance Control of Autonomous Systems

这项教师早期职业发展（Career）计划将为下一代高性能自主系统的控制建立创新的数学和计算决策方法。广泛的新兴技术利用了自主性，包括自动驾驶汽车、用于交付的四旋翼飞行器、搜索和救援机器人、用于环境监测的移动传感器、用于仓库和建筑物的气候控制系统、自动化药物分配和下一代电力网络。自动系统的大规模部署是一项关键的技术飞跃，将从根本上改变我们的生活。自主系统将最大限度地提高效率和可靠性。此外，它们将通过减少或消除人类对危险任务的直接参与来提高安全性。然而，自主系统提出了从根本上困难的技术挑战，以超越当前自动化技术的限制。该项目为一种革命性的自主方法奠定了数学和计算基础，使以前不切实际的计算变得容易处理，并提供了有保证的性能和稳定性水平。为了满足全新的、极具挑战性的性能和可靠性要求，未来的自主系统必须在没有人工操作的情况下做出最佳决策来控制其动作。他们必须充分利用他们的性能信封，同时满足关键任务和环境限制。因此，数学优化问题在自主控制中是普遍存在的。虽然优化提供了一个强大的制定框架，但由于当前算法能力的不足，到目前为止，基于实时优化的控制尚未过渡到普遍实践。为了实现基于实时优化的控制，这项CAREER资助将开发一个全面的凸化理论——将控制问题表述为可处理的凸优化问题——实现目前在许多应用中无法实现的高性能控制。具体来说，它将发展：作为凸优化的自治控制问题的精确公式；定制的数值优化算法，利用问题结构进行极其快速和可靠的机载计算；并采用严格的验证方法来验证所得到的自主控制算法的性能和鲁棒性。这些进步和由此产生的见解也将有利于自主系统的系统工程和设计、高级自主任务规划和自主多主体系统的控制。总的来说，理论和算法产品将形成一个强大的技术基础，允许将基于实时优化的自治系统控制转变为实践。

This Faculty Early Career Development (CAREER) Program grant will build innovative mathematical and computational methods of decision making for control of next generation high-performance autonomous systems. A broad range of emerging technologies utilize autonomy, including self-driving vehicles, quadrotors for delivery, search and rescue robots, mobile sensors for environmental monitoring, climate control systems for warehouses and buildings, automated drug dispensing, and next-generation power networks. Large-scale deployment of autonomous systems is a critical technological leap that will fundamentally transform our lives. Autonomous systems will maximize efficiency and reliability. Furthermore, they will improve safety by reducing or removing direct human involvement in hazardous tasks. However, autonomous systems present fundamentally difficult technical challenges to go beyond the limitations of the current state of the art in automation. This project lays the mathematical and computational foundations for a transformative approach to autonomy, one that makes previously impractical calculations tractable, and provides guaranteed levels of performance and stability. To meet fundamentally new and distinctly challenging performance and reliability requirements, future autonomous systems must make best possible decisions to control their actions without a human operator in the loop. They must utilize their full performance envelopes, while simultaneously satisfying critical mission and environmental constraints. Consequently, mathematical optimization problems are ubiquitous in autonomous control. Though optimization provides a powerful formulation framework, thus far real-time optimization based control has not transitioned to common practice due to shortcomings in the current algorithmic capabilities. To enable real-time optimization based control, this CAREER grant will develop a comprehensive theory of convexification -- formulation of control problems as tractable convex optimization problems -- enabling high performance control that is currently not achievable in many applications. Specifically, it will develop: accurate formulations of autonomous control problems as convex optimizations; customized numerical optimization algorithms that exploit problem structure for extremely fast and reliable onboard computations; and rigorous verification methods to certify the performance and robustness of the resulting autonomous control algorithms. These advances and resulting insights will also benefit systematic systems engineering and design for autonomous systems, high-level autonomous mission planning, and control of autonomous multi-agent systems. Collectively the theoretical and algorithmic products will form a strong technical foundation allowing the transition of real-time optimization-based control for autonomous systems into practice.