Reinforcement-learning-based path planning for UAVs in intensive obstacle environment

In intensive obstacle environment, the available flying space is narrow, which makes it difficult to generate feasible path for UAVs within limited runtime. In this paper, a Q-learning-based planning algorithm is presented to improve the efficiency of single UAV path planning in intensive obstacle environment. By constructing the space-action state offline learning planning architecture, the proposed method realizes the rapid path planning of UAV, and solves the high time-consuming problem of reinforcement learning online path planning. Considering the time-consuming problem of Q-table re-training, a probabilistic local update mechanism is proposed by updating the Q-value of the states to reduce the high time-consuming of Q-table re-raining and realize the rapid update of Q-table. The probability of Q-value updating is up to the distance to the new obstacle. The closer the state is to the new obstacle, the higher its probability of re-training. Therefore, the flight trajectory can be quickly re-planned when the environment changes. Simulation results show that the proposed Q-learning-based planning algorithm can generate path for UAV from random start position and avoid the obstacles. Compared with the classical A* algorithm, the path planning time based on the trained Q table can be reduced from second to millisecond, which significantly improves the efficiency of path planning.