Target Inference Three-Dimensional Intercept-Angle Guidance With Deep Latent Reinforcement Meta-Learning

In this paper, a novel guidance law is proposed for interception of various types of targets at desired impact angle in three-dimensional coupled dynamics. Impact angles are defined using velocity vectors of missile and target, rather than line-of-sight (LOS) angles in azimuth and elevation planes. The proposed guidance law is derived using deep latent reinforcement meta-learning. Different from existing online gradient descent-based reinforcement meta-learning method, the proposed approach is aware of environment variations through latent context vector inferred with amortized variational inference. Target acceleration is also estimated with the latent context vector to facilitate guidance law design. In addition, the proposed approach is built as an adaptive guidance law with range-adaptive hyperbolic tangent function to attenuate control effort. Numerical simulation results with different initial condition and measurement noise validate the effectiveness and robustness of proposed guidance law.