Teaching UAVs to Race Using UE4Sim
By Matthias Mueller, Vincent Casser, Neil G. Smith, Dominik Ludewig Michels, Bernard Ghanem. 2017
10 citations recorded by the source corpus.
Automating the navigation of unmanned aerial vehicles (UAVs) in diverse scenarios has gained much attention in the recent years. However, teaching UAVs to fly in challenging environments remains an unsolved problem, mainly due to the lack of data for training. In this paper, we develop a photo-realistic simulator that can afford the generation of large amounts of training data (both images rendered from the UAV camera and its controls) to teach a UAV to autonomously race through challenging tracks. We train a deep neural network to predict UAV controls from raw image data for the task of autonomous UAV racing. Training is done through imitation learning enabled by data augmentation to allow for the correction of navigation mistakes. Extensive experiments demonstrate that our trained network (when sufficient data augmentation is used) outperforms state-of-the-art methods and flies more consistently than many human pilots.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
- Vladlen Koltun
- Sergey Levine
- Valts Blukis
- René Ranftl
- Antonio Loquercio
- Jitendra Malik
- Alexey Dosovitskiy
- Percy Liang
Citations in the corpus
- Beauty and the Beast: Optimal Methods Meet Learning for Drone Racing (2018, background): “Output representations range from predicting discrete navigation commands (classification in action space) [11], [7], [15] to direct regression of control signals [19].”
- Deep Drone Racing: Learning Agile Flight in Dynamic Environments (2018, methodology): “In contrast to previous work [21], we believe that end-to-end learning of low-level controls is suboptimal for the task of drone navigation when operating in the real world.”
- Following High-level Navigation Instructions on a Simulated Quadcopter with Imitation Learning (2018, background): “Neural network architectures have achieved remarkable performance in various high-level tasks, but their applications to the robotics domain have largely been limited to single, repeated tasks [10, 26, 36] or required substantial amount of training data due to high sample-complexity [39].”