Towards robotic cognition using deep neural network applied in a goalkeeper robot

Developing a vision system combined with a decision system for a humanoid robot, capable of playing soccer in the RoboCup domain, has been proved to be a challenging task. The computational limitations imposed by a embedded computer inside the robot and special conditions, such as the use of colored objects, led teams to use techniques based on color segmentation for vision and conditional statements for decision. However, the current league trend is to insert the robots into more and more realistic environments. This will require the robot to, given an image provided by its camera, to abstract all the information it needs to make a decision regardless of the environment. Most robotic vision systems at RoboCup relies on traditional computer vision techniques: thresholding; windowing; segmentation; and classification that requires hours of labeling to training and testing. This paper proposes a system that does not require to locate objects coordinates in the image - a deep neural network will identify most important features resulting as an output that is a decision. Results show that Deep Neural Network (DNNs) enabled the system to be more simple, robust (with less parameters to be set by hand) and achieved a performance that is compatible with the dynamics of the humanoid robot soccer. This system was tested in a real robot and simulator.