Explainable Deep Reinforcement Learning for UAV autonomous path planning

Autonomous navigation in unknown environment is still a hard problem for small Unmanned Aerial Vehicles (UAVs). Recently, some neural network-based methods are proposed to tackle this problem, however, the trained network is opaque, non-intuitive and difficult for people to understand, which limits the real-world application. In this paper, a novel explainable deep neural network-based path planner is proposed for quadrotor to fly autonomously in unknown environment. The navigation problem is modelled as a Markov Decision Process (MDP) and the path planner is trained using Deep Reinforcement Learning (DRL) method in simulation environment. To get better understanding of the trained model, a novel model explanation method is proposed based on the feature attribution. Some easy-to-interpret textual and visual explanations are generated to allow end-users to understand what triggered a particular behaviour. Moreover, some global analyses are provided for experts to evaluate and improve the trained network. Finally, real-world flight tests are conducted to illustrate that our path planner trained in the simulation is robust enough to be applied in the real environment directly. (C) 2021 Elsevier Masson SAS. All rights reserved.

Automated summary

This paper proposes a novel explainable deep neural network-based path planner for autonomous flight of quadrotors in unknown environments, trained using Deep Reinforcement Learning method in simulation. A model explanation method based on feature attribution is introduced to provide easy-to-interpret textual and visual explanations for end-users to understand the behavior triggers.