Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
End-to-End Learned Vision-Based Navigation
Human pilots are very good at drone racing, even though they only use images as input. In this project, we aim to develop a neural network that can be used to autonomously fly through a sequence of gates while using the same visual information that is also provided to human pilots.
Humans can pilot drones at speeds over 15 m/s through narrow racecourses while only relying on onboard vision. Although humans get better over time, a skilled human pilot will be able to fly through a new course the first time he sees it. For a drone to be able to do the same thing, It
a) needs to identify a gate autonomously
b) fly through the detected gate.
For a) existing approaches can be used that reliably detect a gate. Therefore, the focus of this project is to accomplish item b) using a neural network that operates on the gate detections as an input. The network should not need to be trained on a specific track but rather generalize to new, unseen track-layouts just like the human counterparts.
Requirements:
Machine learning experience (TensorFlow and/or PyTorch), Programming experience in C++ and Python
Humans can pilot drones at speeds over 15 m/s through narrow racecourses while only relying on onboard vision. Although humans get better over time, a skilled human pilot will be able to fly through a new course the first time he sees it. For a drone to be able to do the same thing, It a) needs to identify a gate autonomously b) fly through the detected gate.
For a) existing approaches can be used that reliably detect a gate. Therefore, the focus of this project is to accomplish item b) using a neural network that operates on the gate detections as an input. The network should not need to be trained on a specific track but rather generalize to new, unseen track-layouts just like the human counterparts.
Requirements: Machine learning experience (TensorFlow and/or PyTorch), Programming experience in C++ and Python
Develop and deploy (simulation and, optionally, real world) a neural network controller that flies a drone through a sequence of drone-racing gates.
Develop and deploy (simulation and, optionally, real world) a neural network controller that flies a drone through a sequence of drone-racing gates.
Leonard Bauersfeld (bauersfeld AT ifi DOT uzh DOT ch), Yunlong Song (song (at) ifi (dot) uzh (dot) ch)
Leonard Bauersfeld (bauersfeld AT ifi DOT uzh DOT ch), Yunlong Song (song (at) ifi (dot) uzh (dot) ch)