Risk-aware eVTOL path planning for urban fixed-altitude cruise: A multimodal fusion reinforcement learning approach

Abstract

The complexity and uncertainty of urban environments create an urgent need for safety-enhanced path planning in the autonomous navigation of electric vertical take-off and landing (eVTOL) aircraft. However, existing methods often focus solely on binary collision outcomes or suffer from local optima entrapment, failing to guarantee global path safety and thereby increasing the operational risks. To address these challenges, this paper proposes a risk-aware reinforcement learning (RL)-based framework for eVTOL path planning during fixed-altitude cruise. First, we quantitatively characterize potential and real-time risks by constructing prior risk maps and exploiting temporal radar data, respectively. Subsequently, to mitigate heterogeneous risks comprehensively, we develop a multimodal fusion soft actor-critic (SAC) algorithm, termed MOSSAC. This algorithm leverages a specifically designed encoder to efficiently fuse map, observation, and ego-state information, thereby enhancing the agent’s situational awareness of the quantified risk environment. Ultimately, under the guidance of the risk assessment mechanism, the framework achieves near-optimal path safety performance in real-time navigation. Experimental results demonstrate the superior generalization and robustness of the proposed method. Notably, compared to baseline methods, our method achieves a significant safety improvement of up to 10.84% in dynamic scenarios, with a marginal path length increase of less than 2%.