Field Friction Recognition and State Inference in AI Soccer

Abstract

In this paper, we introduce an approach for friction recognition and state inference under environment where field friction changes. We use a robot soccer-based challenge environment, AI Soccer, where two teams of two-wheeled agents compete to win a game of soccer. We modify the AI Soccer simulator to make the friction coefficient between the agent wheels and the field intermittently changes to add more complexity on the challenge. The friction change causes the AI Soccer's agents to run in a slippery environment time to time that the agents need to adapt their controls in order to play the game properly. As the agents are not equipped with sensors usable for friction detection, we develop a friction recognition classifier with multilayer perceptron based on agents' coordinates and their wheel speed signals. However, the classification of the environmental friction would not be sufficient state information for the agents to perform well under the dynamic environment as behaviors of the ball or opponent agents would be affected by the friction as well. Detailed context information based on friction can help AI Soccer players make better decisions. Therefore, we in addition build a clustering module upon the trained multilayer perceptron classifier. clustering resonance network is adopted as the clustering module for the network's incremental and unsupervised learning capability. Through experiments we demonstrate our system can detect friction state of the AI Soccer field and infer further based on the learned features.