We use a gait generator based on the Robotis Darwin-OP software, that is an open loop walk generator.
At the moment we only detect the ball, using the Mobilenet system trained on thousands of ball images.
The team does not execute localization during the games, only during the start of the game, and it is made by dead reckoning.
The decision process is based on state machines, one for each type of player. States are based on the position of the ball in the image, and actions are done to achieve the ball and kick it. The goalkeeper was trained using deep reinforcement learning.
Out team competes in the humanoid league Since 2014, finishing among the 8 best teams in 2016. Our team was one of the first to use NUC computing units, to use carbon fiber in the structural components of the robots, to use large 3D printed parts, among other small contributions to the league.
Object detection under constrained hardware scenarios: a comparative study of reduced convolutional network architectures. IEEE Latim American Robotics Symposium 2019.