3D UAV path planning in unknown environment: A transfer reinforcement learning method based on low-rank adaption

The increasing number of application scenarios necessitate unmanned aerial vehicles to possess the capability of autonomous obstacle avoidance and navigation in unknown environments, representing a crucial direction for its development. Path planning plays a crucial role in this process. Path planning aims to design efficient and safe navigation paths for UAVs, thereby significantly reducing energy consumption and time spent while improving equipment adaptability to the environment. Firstly, we employ the deep reinforcement learning algorithm to train the agent on randomly changing maps, enabling it to possess both generalization capabilities and active obstacle avoidance skills. Secondly, a novel framework combining transfer reinforcement learning is proposed. It establishes the pre-trained model and utilizes the enhanced low-rank adaptive algorithm to transfer it into formal training, thereby incorporating prior knowledge and improving the efficacy of formal training. Finally, a novel method of sample abundance is proposed to reuse the experience pool stored in the pre-trained model and further increase the generalization capability of the agent, thereby significantly improving its success rate. The proposed algorithm efficiently uses both the pre-trained model and the experience pool. In practical applications, the pre-trained model can be acquired by training on a limited dataset to endow the agent with autonomous obstacle avoidance capabilities. In formal training, numerous random samples are established to simulate unfamiliar environmental terrains. After rapid training, the agent achieves a success rate of 95% in the test set and demonstrates exceptional performance in smoothness and path length.