Invariant Perception for Grasping an Unknown Object Using 3D Depth Sensor

This paper describes robot perception to grasp an unknown object intuitively. Generally, a robot should recognize object´s property such as size, posture and position to make a decision of a grasping task. However, an accurate object information is difficult to perceive quickly from limited measured information, when a robot perceive an unknown object. So, a robot is difficult to decide a suitable action. Therefore, we propose a robot perception system to decide a suitable action intuitively from 3D point cloud data from limited information. We propose robot perception system which is composed of an online process able object extraction method and an invariant detection method for making a decision of a grasping action. The object extraction method can extract a part of point cloud corresponding to an object from 3D point cloud data. The invariant detection method for a grasping action is explained by inertia tensor and fuzzy inference. The invariant for a grasping action affords the possibility of action to a robot directly without inference from object´s property such as size, posture and shape. As experimental results, we show that the object extraction reduces a computational cost drastically, but also the accuracy is better, further the robot can detect a relevant information of a grasping behavior directly.