Computer Vision and Prediction

The only physical sensor for our behavior control software is a S-VHS camera that looks at the field from above and outputs a video stream in NTSC format that is captured by a PC. The computer vision system running on the PC finds and tracks the ball and the robots. We used an improved version of the system described in [2]. Unfortunately, the information about the world is extracted with a delay caused by communication, mechanical constraints, and the computer vision system. The feedback (the result) of an action decision is perceived typically between 100ms and 150ms after the decision has been made. This causes problems when robots move fast, producing overshooting or oscillations in movement control. One way to deal with the delay would be to move more slowly, but this is often not desirable. Our approach to solve that problem is to predict the movement of the robots for the next few frames. We feed the last robot positions and orientations (relative to the current robot position and orientation) and action commands to a feed forward neural network that is trained to predict the robot position and orientation for a point in time 130ms away. We train the network with recorded data before the game. The predicted positions and orientations are used for behavior control. This approximately cancels the effects of the delay and allows for fast and exact movement control.