Safety and Stability Guarantees for Control Loops With Deep Learning Perception
Safety and Stability Guarantees for Control Loops With Deep Learning Perception
复制标题
通过深度学习感知控制回路的安全性和稳定性保证
作者：
Matteo Marchi;Jonathan Bunton;B. Gharesifard;P. Tabuada
深度学习目前用于自主系统的感知管道，例如从相机和LiDAR测量中估计系统状态。虽然这种做法是典型的，但对闭环系统的最坏情况行为的硬保证是罕见的。然而，在这封信中，我们利用神经网络近似的最新结果，结合经典的输入到状态稳定性（ISS）属性，并展示如何设计用于状态估计的深度神经网络，以保证最终闭环系统的安全性和稳定性。
Deep learning is currently used in the perception pipeline of autonomous systems, such as when estimating the system state from camera and LiDAR measurements. While this practice is typical, hard guarantees on the worst-case behavior of the closed-loop system are rare. In this letter, however, we leverage recent results on neural network approximation, combined with classical input-to-state stability (ISS) properties, and show how to design deep neural networks for state estimation that guarantee the safety and stability of the resulting closed-loop system.