Learning Predicted Occupancy Map for Risk-Aware MAV Motion Planning in Dynamic Environments

Mapping the surrounding environment through on-board sensors is crucial for micro aerial vehicles (MAVs) navigating in dynamic unknown environments. Nevertheless, the MAVs' flight safety and efficiency are hindered by limited sensing range and possible obstacle occlusions. In this paper, we present a map prediction method that allows for predicting the occupancy status of a large area surrounding the MAV, that is beyond its sensing range and eliminates the occlusion effects, thus enabling more efficient and safer local motion planning. Specifically, an U-net style with ASPP map predicted model is trained for map prediction and completion about unknown and covered parts. Then based on the more complete map, a risk-aware MAV motion planing framework is presented. We compare our method with several state-of-the-art methods in the PX4+Gazebo simulation environments and perform real-world experiments. The results indicate that our method can achieve a 15% higher collision avoidance rate and a 30% lower navigation stop rate in various environments with indoor and outdoor obstacles.