A Dynamics Predictor-Augmented Soft Actor-Critic for Attitude and Trajectory Control of Unmanned Airships

The tracking of the attitude and trajectory of stratospheric airships is a key technology in aerial intelligent systems, but it presents significant challenges due to complex nonlinear dynamics and susceptibility to environmental disturbances. While deep reinforcement learning(DRL) offers a powerful data-driven approach, model-free methods often suffer from high sample complexity, limiting their practical application. To address this, this paper proposes a dynamics-predictor-augmented Soft Actor-Critic (SAC) algorithm, a hybrid model-based DRL framework for robust airship control. Our approach enhances the highly sample-efficient, off-policy SAC algorithm by integrating a traditional 6-DOF (six-degree-of-freedom) airship dynamics model. This model serves as a predictor to provide the agent with additional information about the system’s expected future states, which stabilizes and accelerates the learning process. A comprehensive simulation environment is established to train and validate the control policy, with the agent’s goal being to achieve high-precision trajectory tracking and stable attitude maintenance. We evaluate our proposed method against a strong model-free proximal policy optimization (PPO) baseline. Experimental results demonstrate that the predictor-augmented SAC controller achieves superior performance, exhibiting significantly higher sample efficiency and learning stability. Consequently, this work offers an effective and scalable solution for the intelligent control of unmanned airships under dynamic uncertainties, highlighting the benefits of combining classical dynamics models with modern deep reinforcement learning techniques.