Abstract
There have been many researchers studying how to enable unmanned aerial vehicles (UAVs) to navigate in complex and natural environments autonomously. In this paper, we develop an imitation learning framework and use it to train navigation policies for the UAV flying inside complex and GPS-denied riverine environments. The UAV relies on a forward-pointing camera to perform reactive maneuvers and navigate itself in 2D space by adapting the heading. We compare the performance of a linear regression-based controller, an end-to-end neural network controller and a variational autoencoder (VAE)-based controller trained with data aggregation method in the simulation environments. The results show that the VAE-based controller outperforms the other two controllers in both training and testing environments and is able to navigate the UAV with a longer traveling distance and a lower intervention rate from the pilots.
Highlights
Multirotor unmanned aerial vehicles (UAV) have achieved considerable success in the past few years due to their high maneuverability and vertical take-off and landing capabilities
It’s noted that while in this paper, we focus on the problem of UAV navigation in riverine environments, the proposed method can be applied to other complex environments which require vision-based solutions and human demonstrations can be collected to benefit the training and allows the agent to learn a good behavior
The linear regressionbased controller achieves a better performance compared to the end-to-end neural network controller but still crashes earlier than a variational autoencoder (VAE)-based controller
Summary
Multirotor unmanned aerial vehicles (UAV) have achieved considerable success in the past few years due to their high maneuverability and vertical take-off and landing capabilities. Autonomous navigation of UAVs with obstacle avoidance capability in outdoor environments remains a challenge especially when the environment becomes complex and unknown (e.g., GPS-denied riverine environments involve heavy foilage/forest canopy, see Fig. 1). Traditional approaches to navigating a robot in complex and GPS-denied environments usually integrate the visualinertial odometry (VIO) or simultaneous localization and mapping (SLAM) techniques [10, 50] with trajectory planning. The optimal trajectory is mapped to the control commands of the robot in order to reach the goal points as well as avoid collisions. While these planning-based approaches have been widely adopted by many scholars [11, 57], the algorithm itself can be very computationally intensive and does not guarantee the performance when environment is non-static. The proposed controller is able to provide feasible solutions to the UAV navigation inside complex riverine environments efficiently and effectively
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
