Neural MPC for quadrotor trajectory tracking on embedded GPUs

关键词:  Unmanned aerial vehicle (UAV) · Model predictive control (MPC) · Real-time trajectory tracking · Deep neural network approximation

Key words:  Unmanned aerial vehicle (UAV) · Model predictive control (MPC) · Real-time trajectory tracking · Deep neural network approximation

摘要:

This paper investigates the computational challenges associated with Model Predictive Control (MPC) for trajectory tracking of quadrotor Unmanned Aerial Vehicles (UAVs). Despite its superior control performance, MPC is difficult to implement in real-time applications on resource-limited embedded systems due to its high computational complexity. Prior research has introduced explicit MPC methods to accelerate MPC computation, but these methods are generally ineffective in embedded applications involving complex systems, such as UAVs, which require high-frequency, real-time control. To overcome these limitations, this study proposes a novel MPC framework for quadrotor UAV trajectory tracking, specifically tailored for embedded GPUs. The method integrates Lyapunov Guidance Vectors (LGV) and Deep Neural Networks (DNN) to significantly reduce computation time. Our approach reduces the computation time to about one thirty-fifth of that required by conventional MPC methods, while still maintaining accurate trajectory tracking performance. Additionally, we validate the effectiveness of the proposed method through Hardware-in-the-Loop (HIL) simulations conducted on an embedded GPU platform (Nvidia Jetson Nano). Numerical results demonstrate that the proposed algorithm reduces the computation time of a high-performance trajectory trackingMPCto 2.98milliseconds, thus providing a promising solution for real-time autonomous control of quadrotor UAVs on embedded systems.

Abstract:

This paper investigates the computational challenges associated with Model Predictive Control (MPC) for trajectory tracking of quadrotor Unmanned Aerial Vehicles (UAVs). Despite its superior control performance, MPC is difficult to implement in real-time applications on resource-limited embedded systems due to its high computational complexity. Prior research has introduced explicit MPC methods to accelerate MPC computation, but these methods are generally ineffective in embedded applications involving complex systems, such as UAVs, which require high-frequency, real-time control. To overcome these limitations, this study proposes a novel MPC framework for quadrotor UAV trajectory tracking, specifically tailored for embedded GPUs. The method integrates Lyapunov Guidance Vectors (LGV) and Deep Neural Networks (DNN) to significantly reduce computation time. Our approach reduces the computation time to about one thirty-fifth of that required by conventional MPC methods, while still maintaining accurate trajectory tracking performance. Additionally, we validate the effectiveness of the proposed method through Hardware-in-the-Loop (HIL) simulations conducted on an embedded GPU platform (Nvidia Jetson Nano). Numerical results demonstrate that the proposed algorithm reduces the computation time of a high-performance trajectory trackingMPCto 2.98milliseconds, thus providing a promising solution for real-time autonomous control of quadrotor UAVs on embedded systems.