Analysis of Illumination Robustness in Long-Term Object Learning

In this article we evaluate the incremental object learning approach of the iCub humanoid robot which is directed towards long-term engagement. Affordable robot companion systems are currently entering the consumer market which highlights the importance in understanding environmental influences on robotic systems under real world conditions. If a robot is to be sent into the real world or different robots/sensors are to be used, we need our algorithms to be independent from both illumination and sensor influenced changes. In our work, we investigate the robustness of the interactive object learning to linear and non-linear lighting changes which can occur due to illumination changes throughout the day or the sensors used. Our results with the models we use suggest that the current method is susceptible to these changes. Therefore, we provide an adjustment to the current method to be able to cope with this problem.