Robots need candidate area where unknown objects exit to find them. We have developed an environmental monitoring sensor which consists of a 3-D LIDAR and a fish-eye camera. It can observe whole room and detect changed areas to gather fine information.

Human finds and recognizes unknown object by using visual clue, tactile, and motion clues. We consider that motion and tactile clues can allow robots to recognize unknown objects. We have developed a method of unknown object recognition and modeling based on visual and motion clues. Grasp-less motion was used to generate the motion.

The recognition of transparent object is one of the important research topics in robotics vision because it is hard to find transparent objects by using visual clues. Robot need to find and grasp them in real environments. We proposed a method of transparent object detection by using range sensors. Distortion caused by transparent objects is used for detecting them.

People use bookshelves, cupboards, and racks for cleaning up. Usually similar objects are gathered and piled at same location, however it is hard for robots to pick up an object from the piled objects. In this research, we aim to find target object from the piled object and grasp it. Segmentation is a key research topic. Visual feature is not sufficient but Grasp-less manipulation is efficient enough for segmentation of the piled objects. We propose a method of object segmentation by using visual and motion clues.

Partial observation Markov decision process (POMDP) can allow us to decide suitable motion for unknown object segmentation. When robots grasp unknown object, the robots treat uncertainty of motion and recognition. POMDP can treat such uncertainties and decide the suitable motion based on prior experience. We proposed a motion planning method based on POMDP which can select suitable motion on the basis of geometrical features.