Comparison Between Two Recurrent Multitasking Networks for Robot Relocation

• Dunfrey Pires Aragão UFRN
• Tiago Nascimento UFPB

Resumo

Perform multitasking solutions has shown a degree of effort because, while making it available to achieve a new task, it impairs the previous accuracy and increases the errors computed to come up by the single-task method. A possible case associated with these increasing errors may is related to the association with the contextual dependency that exists on temporal depending tasks. Recurrent neural networks are excellent approaches to approximate states such as the pose in mapping exploration, a problem that has temporal data dependence. Mapping exploration has three important goals: the agent’s (robot) location on the map, its orientation, and depth-scene discernment of objects in the view. In this work, we present a model using two state-of-the-art neural networks, GoogLeNet and U-Net, applied to LSTM layers, to achieve three goals simultaneously: the pose’s (location and orientation) agent, and depth-scene image, by given a single RGB image as input data, evaluated with the ContextualNet to comparing the accuracy in pose regression, based on two datasets on robot experiments.

Palavras-chave: Training, Recurrent neural networks, Robot vision systems, Pose estimation, Predictive models, Multitasking, Data models