Deep ART memory based cognitive architecture for robots

Abstract

Robots have been expected to do various troublesome or hard works for humans. Since these human-scale tasks consist of sequential procedures, a memory structure storing the sequences effectively is essential for robots to perform tasks. In this paper, we propose the new memory structure, called Deep ART, to learn time sequential events of the tasks from user's demonstration. After learning the tasks, it can retrieve a proper episode related with the current situation from the image input. In contrast with previous memory models, the retrieval process is more robust from noisy inputs or partial cues since its different encoding and decoding processes. Moreover, a cognitive architecture based on Deep ART is designed for robots to perform tasks in the human environment. Our proposed architecture has the perception module of which the function is image and voice recognition, and the episode inference module that enables robots to infer the appropriate episodes from user's natural language commands. The experimental results show the effectiveness of our architecture in terms of learning and executing tasks, and our robot successfully completed multiple tasks related with user's voice command.