G&CNets - On-Board Optimal Control of a Quadcopter
In collaboration with TU Delft, we have further investigated the use of G&CNets for real-time optimal control by training and testing these networks onboard the Parrot Bebop drone in the TU Delft MAVLab.
By training a deep neural network on the optimal controls for time optimal flight, we were able to outperform one of the more commonly used, state-of-the-art controllers based on minimising the snap (4th time derivative of position). Below are videos showing the flight path and timing of the two controllers starting from the same position.
Benchmark controller:
G&CNet Controller:
These videos show that the G&CNet controller follows a similar control pattern, but significantly more aggressive for faster flight. All of the control computation is done on-board based on the state of the quadcopter thus showing that it is possible to run a neural network on-board and achieve optimal flight.
A publication has been submitted to IEEE Robotics and Automation Letters and is awaiting review. Here you can find our previous work on G&CNets, and here the application of G&CNets to interplanetary trajectories. This page will be updated when the paper is published so stay tuned!