Prior-controller-assisted multi-agent deep reinforcement learning attitude control for hypersonic morphing vehicles

Hypersonic morphing vehicles have attracted significant attention due to their large planform configurations and wide-speed-range flight capabilities. However, the morphing process induces rapid changes in aerodynamic parameters and introduces pronounced system uncertainties, which considerably complicates attitude control. To address this issue, a prior controller-assisted multi-agent deep reinforcement learning method was developed for the attitude control of hypersonic morphing vehicles. The control problem was reformulated as a cooperative decision-making task with three independent agents, each assigned to an individual attitude channel, thereby improving control performance and decision stability. A prior controller-assisted safe experience collection strategy was introduced, in which prior controllers were incorporated during training to guide attitude regulation and suppress random exploration in divergence-prone regions. An adaptive trust mechanism based on a Critic network was designed to evaluate intelligent control performance and dynamically regulate the selection probability between intelligent and prior-controller-based actions, enabling a smooth transition from prior controller-assisted guidance to purely intelligent exploration. The simulation results demonstrated that the proposed method reduced attitude divergence events during the training phase from 1500 to 31, highlighting a significant improvement in training stability and exploration safety, increased cumulative rewards by 3.5% compared with the Soft Actor-Critic (SAC) algorithm, and improved overall control performance by 10.68%.