A Lunar Robot Obstacle Avoidance Planning Method Using Deep Reinforcement Learning for Data Fusion

In future exploration and base construction on the moon, obstacle avoidance planning of lunar robots in an uncertain environment is critical for their autonomous movements and operations, with no precise location information of obstacles. In the present work, an obstacle avoidance planning method using deep reinforcement learning with a double-channel Q network is proposed, by which local surveillance video images and navigating data are merged for action value estimation. Through simulation, our method is turned out to achieve motion planning effectively from raw sensing data, and learn faster than the methods using single type of data.