Design of Object Learning System by Using Random Forest, Convolutional Neural Network and Gated Recurrent Neural Network for Service Robot

• Main Authors: Cheng-HuiLi, 李政輝
• Other Authors: Tzuu-Hseng Steve Li
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/nvz2ah

碩士 === 國立成功大學 === 電機工程學系 === 105 === Inspired by the self-exploring learning approach, this thesis proposes an object learning system in which the robot interacts with an object to obtain its features and constructs the object concept. The system consists of three kinds of features: interaction features, visual features, and intrinsic features. When the robot interacts with an object, for example pushing or stacking the object, it observes the changes of the object to obtain the interaction features. At the same time, the robot gets the visual features of the object that include color image information and depth information. The intrinsic features are the properties of an object, such as volume, weight and shape. The relationship models of these three kinds of features are constructed through a Random Forest algorithm (RF) and a convolutional neural network (CNN). The established models help the robot to predict the properties of a new object and to make decisions. The relationship models between the interaction features and the intrinsic features are built by the RF, where an artificial bee colony algorithm is integrated as the splitting function of each node and then used to judge whether a feature is good or not. The relationship models between the visual features and the intrinsic features are determined by the CNN, which allows the robot to decide how to interact with an object through the obtained visual features. Two experiments are constructed in this thesis, the service providing task and the stacking task. In the former, the robot figures out the appropriate object by the object concept models to accomplish the appointed task, for example, selecting a suitable container to pour water. In the latter experiment, the robot combined with the gated recurrent neural network to learn out the stacking sequence of various objects. All the real experimental results demonstrate that the robot can build the object concept models by interacting with the objects and utilize these models to accomplish several tasks.