Evaluating Data Collection Methods for Vision-Based Learning in Humanoid Robot Soccer

摘要

Humanoid robots competing in RoboCup (an international robot soccer competition) must perceive their environments under highly dynamic and often unpredictable conditions. Requiring a vision system for localization and path planning, teams typically need to collect training image data for the machine learning object detection model. This paper presents an empirical comparison of three commonly used methods - handheld camera, gimbal-mounted systems, and rollable tripods. Experimental results show that the gimbal-mounted approach consistently outperforms the other two, yielding superior precision and recall when detecting crucial soccer field landmarks and the game ball. These results highlight that data collection methods which effectively simulate the robot's actual visual experience during gameplay lead to more robust and reliable vision models. Inspired by these findings, we implemented a revised vision pipeline in our latest humanoid robot, ARTEMIS, capturing data directly from its onboard stereo camera system. This approach proved instrumental in achieving reliable object detection in real-time, even under severe motion blur and degraded field conditions during the dynamic matches, resulting in our eventual victory. We discuss the advantage and limitations of each data collection method, emphasizing the critical role of matching the robot's real-world visual experience to achieve champion-level performance.