Integration of vision and walking skills together with high-level strategies for quadruped robots to play soccer

Abstract

Legged robots taking part in real multi-agent activities represent a very innovative challenge. This domain of research requires developments in three main areas. First, without any feedback information from the environment, there is no way for robots to achieve some tasks autonomously. Fortunately, the quadruped 'Sony' prototypes on which all experiments are carried out are equipped with an enhanced vision system; thanks to its CCD camera located in its head, the robot can obtain color images of the scene around it. Extracting relevant information from the images captured is not easy since it must be done onboard in real time. Moreover, image treatment procedures should have high process rates for the robot to react quickly in front of unexpected events. A special vision module composed of three parts has been designed for these purposes. The second point to focus on is the walking ability of the robot. Quadrupeds are designed to move efficiently and rapidly on flat ground. The objective of the walking module is to generate appropriate walking patterns allowing the machine to walk in the desired direction. Walking gaits are produced like reflexes by the robot itself to adapt to the situation. With regard to the design of these gaits, emphasis has been put on increasing speed and mastering transitions. Finally, the machine should be given a minimum of intelligence since it has to manage vision information and its walking gaits by itself. When involved in situations of cooperation or competition or both, like in a soccer game, a high-level supervision task is welcome. This paper presents detailed developments of these three points and describes how they are implemented on a real robot.