Extremum-seeking active object recognition in clutter using topological descriptors

Object recognition in unseen and cluttered indoor scenes is a challenging problem for semantic-level mapping and manipulation tasks involving low-cost mobile robots. In this paper, we propose a novel framework to address this problem through active robot navigation. Using this framework, the robot performs instance segmentation and identifies the objects using a 3D point cloud slicing-based topological descriptor. It also optimizes its pose autonomously via an extremum seeking controller to improve the identification confidence scores. Results show that our framework always improves the recognition success rate for any given scene as the robot moves to better pose(s), regardless of the number of objects in the scene, degree of clutter, distance to the objects, and lighting condition.