Add-if-Silent Rule-Based Growing Neural Gas for High-Density Topological Structure of Unknown Objects
Masaya Shoji, Takenori Obo, Naoyuki Kubota · 2023
To realize a super-smart society (Society 5.0) where humans and robots coexist, there is a need for a perceptual system that can recognize unknown objects in various unknown environments quickly and flexibly. In unknown environments, the characteristics of objects cannot be known in advance, so prior learning-based recognition methods such as deep reinforcement learning cannot fully cover the problem. There have been many studies on environment recognition (clustering, etc.) using a combination of RGB images and distance images, but the recognition performance is unstable because it strongly depends on the lighting conditions of the environment. Therefore, in this study, we construct a 3D topological map of the environment in real-time using Growing Neural Gas (GNG), which can learn 3D topological structures even for unlearned objects, using only 3D point cloud data as input. In the real world, due to the characteristics of RGB-D cameras, sample density decreases for more far-away objects and only sparse depth information can be obtained, so conventional GNG cannot generate high-density topological structures of unknown objects. Therefore, if the object category labels of the winner nodes (nearest nodes) for the input vector (3D point cloud) match the unknown object and are within a predefined tolerance area, then it is judged to be useful input information for learning the topological structure of the unknown object, and the topological structure of the unknown object is learned. We propose Add-if-Silent rule-based GNG (AiS-GNG) which can generate high-density topological structures for far-away objects by directly adding input data as a reference vector. We verify the effectiveness of the proposed method through experiments using a 3D dynamics simulator.