Deep Reinforcement Learning‐Based Integrated Guidance and Control With Impact Angle and Field‐of‐View Constraints Under Bearing‐Only Measurements

ABSTRACT

Increasingly complex combat missions impose higher demands on missile guidance performance, making the development of strapdown miniature missiles an important trend. Addressing the integrated guidance and control (IGC) problem for strapdown miniature missiles with impact angle and field of view (FOV) constraints under bearing‐only measurements, this paper proposes a deep reinforcement learning (DRL)‐based IGC approach. The missile's kinematic and dynamic models are formulated as a discrete‐time Markov decision process (MDP). The impact‐angle error and body‐line‐of‐sight (BLOS) angle are selected as the state variables, and a corresponding reward function is constructed. Policy and value networks are established accordingly, with fin deflection serving as the policy action. Building on proximal policy optimization (PPO), this paper introduces a segment‐wise advantage sampling strategy and a sliding‐window data update method. An elite trajectory buffer is maintained to guide training, and the training data are refreshed in a sliding‐window manner. The training results indicate that the improved algorithm provides more stable and efficient learning under the tested conditions. Policy test results indicate that the proposed IGC approach maintains excellent guidance performance under varying conditions, exhibiting strong robustness and a capture region comparable to or larger than those of existing guidance laws while using less measurement information. These results validate the effectiveness and superiority of the presented method.