Active Object Categorization on a Humanoid Robot

We present a Bag of Words-based active object categorization technique implemented and tested on a humanoid robot. The robot is trained to categorize objects that are handed to it by a human operator. The robot uses hand and head motions to actively acquire a number of different views. A view planning scheme using entropy minimization reduces the number of views needed to achieve a valid decision. Categorization results are significantly improved by active elimination of background features using robot arm motion. Our experiments cover both, categorization when the object is handed to the robot in a fixed pose at training and testing, and object pose independent categorization. Results on a 4-class object database demonstrate the classification efficiency, a significant gain from multi-view compared to single-view classification, and the advantage of view planning. We conclude that humanoid robotic systems can be successfully applied to actively categorize objects a task with many potential applications ranging from edutainment to active surveillance.