Thesis Presentation

Visual object recognition, world state estimation, and localization are challenging problems for a mobile robot. The problems are especially difficult on the AIBO robot platform, which must rely only on local sensor information retrieved from a single monocular camera with limited viewing angle. In the RoboCup domain, specifically the Four-Legged League, teams of autonomous AIBO robots play soccer against one another. This thesis focuses on the creation of intelligent robot soccer behaviors that rely on direct visual input to carry out critical tasks where previous localization-based behaviors have lacked in effectiveness and efficiency. We present novel algorithms for the real-time visual detection of the goals and teammate and opponent robots on the playing field, both of which are capable of detection ranges up to three times greater than in previous implementations, aimed at improving overall team game play. The thesis discusses the ‘shoot on goal’ behavior and the ‘p! ass to teammate’ behavior, which were both modified to make use of direct visual input from the new vision algorithms for improved performance. Experimental results comparing the effectiveness of the proposed visual input-based behaviors against their traditional localization-based implementations are also provided. The thesis concludes with a discussion of possible future work and how the robot team, equipped with the algorithms presented, has fared during actual games at the international RoboCup competition.

Thesis Committee:
Manuela Veloso, Chair
Paul E. Rybski