Building object models through interactive perception and foveated vision

Autonomous robots that operate in unstructured environments must be able to seamlessly expand their knowledge base. To detect and manipulate previously unknown objects, a robot should be able to acquire new object knowledge even when no prior information about the objects or the environment is available. Additional information that is needed to identify new objects can come through motion cues induced by interactive manipulation. In the proposed system, changes in the scene are caused by a teacher manipulating the object to be learned. We propose to improve visual object learning and recognition by exploiting the advantages of foveated vision. The proposed approach first creates object hypotheses in peripheral stereo cameras. By directing its attention towards the identified object area in the foveal views, the robot can conduct a more thorough investigation of a smaller area of the scene, which is seen in higher resolution. We compare two methods for validating the hypotheses in the foveal views and experimentally show the advantages of foveated vision compared to stereo vision with a fixed field of view.