An End-to-End UAV Obstacle Avoidance Approach Fusing Wavelet Convolution and KAN Networks

Traditional UAV navigation methods typically depend on explicit environmental priors and global path planning algorithms, which become ineffective in the absence of high-precision maps or 3D reconstruction data. Consequently, learning-based end-to-end vision-to-control networks have shown substantial promise for the online control of UAVs in complex environments. To overcome the challenge of autonomous obstacle avoidance and navigation in complex environments, this work presents an end-to-end framework based on deep reinforcement learning, specifically designed for UAVs operating in near-ground flight conditions with uncertain scene characteristics. We employ the Kolmogorov-Arnold Network (KAN) for function approximation within the Actor-Critic (AC) learning framework. Meanwhile, a wavelet convolutional network is adopted to process the raw signals from the airborne depth vision sensor. This architecture enables the model to efficiently and accurately process environmental data, thereby generating adaptive obstacle avoidance strategies. Compared to other baseline approaches, such as feature extractors based on traditional CNNs and ResNet18 combined with AC frameworks utilizing feed-forward networks, our method achieves higher obstacle avoidance success rates and a faster convergence rate.