Domain Adaptation from Public Dataset to Robotic Perception Based on Deep Neural Network

Abstract

A robot needs to understand the environment in a human-like manner in order to provide good service for us. However, the working environment changes from time to time for reasons such as light and layout. As a result, the robot needs to adapt existing knowledge to fit current environment. Deep neural network has shown its advantages in feature extraction which could be used in later processing. In this paper, we use a framework of deep neural network, Mask R-CNN, to endow a robot with the capability of adapting knowledge from public dataset to current environment instead of training from scratch. The effectiveness of our proposed adaptive method is evaluated on an actual humanoid NAO robot's environment, where the robot could intelligently perceive the classification and localization of objects, as well as their spatial layout even they are placed randomly.