Direct Perception and Action Decision for Unknown Object Grasping

Introduction

Recently, various types of social robots have been developed for the next generation society of humankind (Beetz, 2011). For example, rehabilitation robots, education robots and therapy robots will be expected for the progress of society (Robins, 2005). Moreover, amusement robots, service robots and partner robots are expected in order to have a comfortable life. The social robots have to work not only in specific environments like factories, but also in general environments such as public facilities and homes (Mitsunaga, 2008). In a general environment, a social robot should take action flexibly in order to fulfill specific tasks, even in an unknown environment. A social robot needs human like perception and action decision in unknown environment. To discuss social robots, there are important consideration such as embodiment, social interaction, perception, interpretation, communication, experience, learning and so on (Fong, 2003). Especially, Wainer et al. insists that robots with physical bodies has a measurable effect on performance and perception of social interaction (Wainer, 2006). A perception of social robots cannot separate with the embodiment of a robot. Therefore, we focus on the discussion of the perception of a robot which has an embodiment to realize a social robot working in an unknown environment.

There are many previous researches that focus on the unknown object recognition method in a real environment (Bay, 2008; Comaniciu, 2001). To perceive an unknown object, F. Jurie et al. had proposed the real-time 3D template matching to detect objects in 3D space (Jurie, 2001). However, this method needs a minimal level of predefined knowledge which is given by the operator, or achieved by oneself. Therefore, a robot cannot perceive the environment and take suitable action immediately in an unexpected situation. All of the 3D template data makes nearly impossible to provide beforehand by human operator in real environment. On the other, there are segmentation methods for detecting an unknown object. Random sample consensus (RANSAC) is a popular segmentation method (Fischler, 1981). RANSAC and progressive RANSAC (PROSAC) (Chum, 2005) are installed in the Point Cloud Library (PCL) (Aldoma, 2012). These methods can detect same features from point cloud data using depth sensor. However, these methodologies require a significant volume of a computational cost to obtain accurate results.