This video shows a quadrocopter learning to fly arbitrarily configured slalom courses in minimum time.
Slalom poles are arbitrarily placed in the space and a time-optimized trajectory around the sticks is calculated. The quadrocopter iteratively learns to follow this pre-calculated trajectory. The vehicle executes its learning iterations above the poles. Once the quadrocopter is guaranteed to pass between the slalom poles, the height of the quadrocopter is lowered and the slalom is performed.
Check out what happens if the quadrocopter performs the slalom without learning!
Our goal is to enable autonomous systems - such as the quadrocopter in the video - to 'learn' the way humans do: through practice. We use the information of previous experiments to improve the vehicle's performance. After each trial we update the input to the vehicle based on the past measurements. We use a model-based update rule that takes into account prior knowledge about the vehicle dynamics and efficiently combines it with experimental data. Such an approach outperforms pure feedback control, since it is able to anticipate recurring disturbances and to actively compensate for them.
Work by Angela Schoellig, Fabian Mueller and Raffaello D'Andrea.
At the ETH Flying Machine Arena, http://www.FlyingMachineArena.org. More information: http://www.idsc.ethz.ch/Research_DAndrea/Iterative_Learning.
This work is supported by and builds upon prior contributions by past and present collaborators at the ETH Flying Machine Arena: http://www.idsc.ethz.ch/Research_DAndrea/FMA/participants.
