The development of high-performance autonomous multi robot control systems requires intensive experimentation in controllable, repeatable, and realistic robot settings. The need for experimentation is even higher in applications where the robots should automatically learn substantial parts of their controllers. We propose to solve such learning tasks as a three step process. First, we learn a simulator of the robots' dynamics. Second, we perform the learning tasks using the learned simulator. Third, we port the learned controller to the real robot and cross validate the performance gains obtained by the learned controllers. In this paper, we describe M-ROSE, our learning simulator, and provide empirical evidence that it is a powerful tool for learning of sophisticated control modules for real robots.

In many multi robot applications, such as robot soccer, robot rescue, and exploration, a reliable coordination of robots is required.

. In this paper we propose a hybrid navigation planning and execution system for performing joint navigation tasks in autonomous robot soccer. The proposed system consists of three components: an arti cial neural network controller, a library of software tools for planning and plan merging, and a decision module that selects the appropriate planning and execution methods in a situation-specic way. The system learns by experimentation predictive models for the performance of dierent navigation planning methods. The decision module uses the learned predictive models to select the most promising planning method for the given navigation task. In extensive experiments using a realistic and accurate robot simulator that has learned the dynamic model of the real robots we show that our navigation system is capable to (1) generate fast and smooth navigation trajectories and (2) outperform the state of the art planning methods. 1

This paper describes the computational model underlying the AGILO autonomous robot soccer team and its implementation. The most salient aspects of the AGILO control software are that it includes (1) a cooperative probabilistic game state estimator working with a simple off-the-shelf camera system; (2) a situated action selection module that makes amble use of experience-based learning and produces coherent team behavior even if inter-robot communication is perturbed; and (3) a playbook executor that can perform preprogrammed complex soccer plays in appropriate situations by employing plan-based control techniques. The use of such sophisticated state estimation and control techniques characterizes the AGILO software. The paper discusses the computational techniques and necessary extensions based on experimental data from the 2001 robot soccer world championship.