Reinforcement Learning-driven automatic calibration for color segmentation-based robot detection

• Clara Acruchi UFPE
• Geovany Icaro L. Nunes UFPE
• Thiago José A. Souza UFPE
• Breno Cavalcanti UFPE
• Adiel Teixeira de A. Filho UFPE
• Mateus G. Machado UFPE
• Edna N. Barros UFPE

Abstract

The systems designed to accomplish the color segmentation task are based on machine-learning-based color classifiers or deterministic algorithms, which usually rely on manual parameter calibration to perform well. Robot soccer competitions such as IEEE VSSS and RoboCup SSL rely on the latter type of vision software and thus suffer from manual calibration, which can be time-consuming and highly depend on operator proficiency and lighting conditions. As a challenge of IEEE VSSS, the teams must develop their robot detection software. Therefore, the present work proposes automatic calibration for color segmentation and robot detection systems as a black-box optimization problem. Then, applying the formulation of the IEEE VSSS robot detection software used in LARC 2023, it compares different methods to solve the problem with multi-armed bandit reinforcement learning algorithms. The analysis includes a detailed evaluation of the performance and efficiency of these methods in enhancing the precision and reliability of robot detection.

Keywords: Software algorithms, Lighting, Color, Reinforcement learning, Manuals, Calibration, Software reliability, Robots, Optimization, Sports, color segmentation, object detection, black box optimization, reinforcement learning, automatic calibration, multi-armed bandit