2025/06/30 by Bin Zhao, Qinglong Zhang, Xiaoyang Huang +2 · 1 citation
Engineering · Computer Science · #Guidance and Control Systems #Target Tracking and Data Fusion in Sensor Networks #Adaptive Control of Nonlinear Systems
paper · doi:10.1109/taes.2025.3584898
In this paper, a three-dimensional (3D) deep reinforcement learning proportional navigation guidance (DRLPNG) law is proposed to intercept high-speed maneuvering target, taking into account the constrained seeker's field-of-view (FOV) and dynamic characteristics of the autopilot. The recommended approach is different from the current guidance laws in that both the passive and active seekers are considered to enhance its adaptability for practical implementation. Firstly, the scenario of intercepting high-speed maneuvering targets is modeled as a standard Markov Decision Process (MDP), incorporating different state transitions according to different seekers. Secondly, a comprehensive reward function is designed to address the sparse rewards during the interception process and ensure strict adherence to the seeker's FOV limit. Thirdly, the guidance strategy that maps observations to the bias term of the proportional navigation guidance (PNG) is produced by employing the twin delayed deep deterministic policy gradient (TD3) algorithm to the interceptor interception model. Simulation results of multiple scenarios demonstrate that the proposed DRLPNG can successfully intercept high-speed maneuvering targets while maintaining strict FOV constraints regardless of the active or passive seekers. The suggested guidance law is user/designer friendly to relax the design effort of the traditional FOV limited guidance law and enhance the adaptability to maneuvering targets.