Robust UAV navigation under perception uncertainty: a spatiotemporal transformer-enhanced reinforcement learning approach

Reinforcement Learning (RL) has become a prominent approach for autonomous unmanned aerial vehicle (UAV) navigation due to its end-to-end nature and low dependence on precise environmental models. However, existing RL methods typically assume ideal sensor perception, whereas real-world perceptual data is often corrupted by noise arising from sensor limitations and environmental factors. To address this challenge, this paper proposes a s patiotemporal t ransformer- e nhanced RL (STERL) method for UAV autonomous navigation. STERL employs dedicated temporal and spatial branch sub-networks to process a temporal feature matrix and a state differential matrix derived from sequential state information. Flight command decisions are then derived by integrating these spatiotemporal representations. Furthermore, by embedding transformer modules within the spatiotemporal dual-branch network to enhance long-term dependency modeling, the method improves navigation safety and flight efficiency while maintaining baseline navigation performance. Extensive experiments demonstrate that the proposed algorithm achieves optimal navigation success rates across varying obstacle densities while exhibiting leading performance in safety and efficiency metrics. Across various noise interference levels, STERL exhibits exceptional robustness, with only an 18.9% performance degradation under extreme noise—significantly lower than existing algorithms. These results highlight its superior performance in complex interference scenarios and strong practical applicability in industrial deployments.