Can deep neural networks also be used to control a robot?

(08-05-2018) If you would like to know more... on May 15th, Jonas Degrave, former IDLab, is defending his PhD on Incorporating Prior Knowledge into Deep Neural Network Controllers of Legged Robots.

 

In the research on AI, deep neural networks are now a big hype. In a short time they managed to teach computers to understand what is reflected in a picture or how a text is put together. This doctoral thesis explores whether these deep neural networks can also be used to control a robot.

For that research, a small four-legged robot dog was developed to verify these methods. Afterwards, a simulation program was developed to simulate the body of the robot, so that both the hardware and the software of the robot could be optimized with an efficient algorithm, based on the analytical derivatives of this system.

The researchers conclude that there can be a natural connection between the physique of the robot and the neural networks that control it. If the robot is built from soft, bending materials, the robot can be controlled with less processing power and less memory, by using the chaotic behavior in the body of the robot.

 

Contact: Jonas Degrave - Jonas.Degrave@UGent.be
Promotors: prof. Francis wyffels & prof. Joni Dambre