Learning-based artificial potential field guidance for hypersonic glide vehicle penetration

Abstract This study investigates the penetration guidance problem for hypersonic glide vehicles in complex constrained environments and proposes a learning-based efficient evasion guidance framework. First, a baseline strategy for determining the vehicle’s reference heading in obstacle environments is designed using an artificial potential field method. This baseline algorithm mitigates the inherent contradictions arising from the direct integration of the artificial potential field into a penetration guidance framework. Second, the key sensitive parameter set affecting penetration performance is identified. Based on a deep reinforcement learning (DRL) framework combining the multi-head self-attention mechanism and twin-delayed deep deterministic policy gradient (TD3), an online parameter optimiser dependent on time-series state data is designed. Meanwhile, to ensure the effectiveness of strategy learning for this optimiser and avoid sparse rewards, three reward function models are constructed to guarantee the stable convergence of the agent’s strategy learning. This optimiser generates real-time responses for evasion strategy adjustments according to environmental dynamics, thus avoiding the potential limitations of manual offline adjustments and the rigidity of online-loaded strategies. Finally, a series of ablation studies and comparative experiments are conducted. These experiments systematically verify the effectiveness and superior performance of the proposed strategy under both standard and disturbed environmental conditions.