A UAV Path Planning Method Based on Deep Reinforcement Learning With Dense Rewards

Most state-of-the-art (SOTA) unmanned aerial vehicle (UAV) path planning approaches depend on global environmental knowledge. While algorithms like adaptive soft actor-critic (ASAC) have improved training efficiency, their obstacle avoidance in partially observable environments remains limited. To address this, we propose a depth-based collision risk prediction (DCRP) algorithm that integrates into the ASAC framework. DCRP processes depth images alongside UAV pose and velocity to calculate a dense collision risk signal, enriching the reward function for more effective avoidance learning. Furthermore, we enhance the policy network with a novel skip connection that directly injects critical state information into the final action output. This innovation mitigates gradient vanishing and accelerates policy learning. Additionally, a generalized transfer learning (GTL) strategy accelerates convergence in complex environments by leveraging policies pre-trained in simpler ones. Extensive evaluation in high-fidelity AirSim environments demonstrates the superiority of our method. It outperforms several SOTA baselines, achieving an approximately 20% higher task success rate and 39% faster training efficiency on average, while maintaining a real-time inference time of around 15 ms.