Deep Learning Methods for Perception and Navigation of Service Robots

Mobile service robots;Deep learning;Reinforcement learning;Computer vision;153

인문대학 협동과정 인지과학전공 ;

University:서울대학교 대학원

关键词: Mobile service robots;    Deep learning;    Reinforcement learning;    Computer vision;    153;   

Others  :  http://s-space.snu.ac.kr/bitstream/10371/142421/1/000000149312.pdf

美国|英语

来源: Seoul National University Open Repository

【 摘 要 】

To assist humans in everyday life, service robots require perception ability to understand the holistic context of the environment by integrating multimodal sensory information. Moreover, service robots have to be able to navigate freely in human environment to actually provide services to humans. Here we propose novel deep learning approaches to improve the perception and navigation performance of service robots. First, we introduce IPSRO (Integrated Perception for Service RObots) framework, which is ROS-friendly integrated perception system that we have recently open-sourced. IPSRO can flexibly integrate several perception modules including state-of-the-art deep learning models to extract rich and useful perceptual information from the environment based on a unified perception representation. On top of that, IPSRO can process the generated perceptual information to perform complex perception tasks. Second, we propose a NavNet, which is a deep reinforcement learning model for robust autonomous navigation. NavNet uses state-of-the-art off-policy deep reinforcement learning model to utilize previous experiences and human demonstration. We present the two system;;s capability evaluated in home-like environments. We also demonstrate the results of the deployment of IPSRO on the RoboCup@Home 2017 Social Standard Platform League.