Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Designing Flying Camera Shots with Deep Learning
The aim of the project is to enable end-users to design robot motions. The focus is on designing motion primitives for cinematography quadrotors.
Keywords: Quadrotors, Control Systems, Deep Learning
Design of robot motions is a tedious process reserved for experts. Depending on the application, the end-user has different requirements for the robot. This results in the expensive process where the end-user and the expert need to work together to achieve the desired solution. To overcome this issue and to accelerate the design process, the learning algorithms are applied directly to the data generated by the end-user. This class of algorithms is called Learning from Demonstration. In our recent work, we show possibility to train dynamic quadrotor motions using path splines as examples. Path splines are intuitive to produce and can be easily produced even by end-users. However, the algorithm requires a lot of example paths. To overcome this issue, the goal is to use Deep Learning techniques to generate training dataset of example paths from a small set of end-user examples. Our focus is on areal cinematography with autonomous drones. The goal is to enable the end-user to fully design cinematography motion primitives.
Design of robot motions is a tedious process reserved for experts. Depending on the application, the end-user has different requirements for the robot. This results in the expensive process where the end-user and the expert need to work together to achieve the desired solution. To overcome this issue and to accelerate the design process, the learning algorithms are applied directly to the data generated by the end-user. This class of algorithms is called Learning from Demonstration. In our recent work, we show possibility to train dynamic quadrotor motions using path splines as examples. Path splines are intuitive to produce and can be easily produced even by end-users. However, the algorithm requires a lot of example paths. To overcome this issue, the goal is to use Deep Learning techniques to generate training dataset of example paths from a small set of end-user examples. Our focus is on areal cinematography with autonomous drones. The goal is to enable the end-user to fully design cinematography motion primitives.