Deep Reinforcement Learning for UAV Indoor Navigation Through Task Decomposition

Dear Editor, This letter addresses the challenges of sparse and delayed rewards in complex indoor navigation tasks. To this end, we propose a task decomposition-based reinforcement learning framework that integrates a reinforcement learning (RL) algorithm with a path planner. Specifically, the rapidly-exploring random tree star (RRT*) algorithm is employed to generate a sequence of sub-goals, which are incorporated into the state space. This decomposition transforms the original long-horizon task into a series of easier sub-tasks with reward monotonicity, providing valuable spatial priors for the unmanned aerial vehicles (UAVs) and guiding it toward more effective exploration. As a result, the proposed method enhances learning stability and mitigates the negative effects of sparse and delayed rewards, facilitating the learning of an optimal navigation policy.