Some of main problems in robotics are compute power and knowedge sharing limitations. There
is still a big gap between the abilities of individual robots and what is required for truly useful au-
tonomous service robots. Network robotics can help to solve the problem by allowing individual
robots to share and pool resources. However, large-scale distributed robot learning issues still need
to be resolved. In this dissertation, I will address the research problem of enabling large-scale dis-
tributed robot learning through cloud computing, with a specific focus on elder care assistance in
fetching small objects. I propose to develop a new service robot infrastructure based on cloud com-
puting. I will design and implement private cloud infrastructure to support large-scale distributed
robot learning to help improve grasping strategies for individual robots.
I will develop the private cloud sharing of infrastructure for cloud robotics on an OpenStack (2013)
and Hadoop (2013) computer cluster. This cloud will support computational power and knowledge
of grasping strategies. I will design and implement object recognition and grasping point detection
for fetching small objects such as eyeglasses, pens, and bottles.
The overall architecture of my proposed methodology is presented in Figure 3.1. I propose cloud sup-
ported large-scale distributed learning based on lightweight robots. I assume the robots participating
in the distributed system will have a single Kinect and communicate over Wi-Fi. The infrastructure
system will consist of five main components: common interface, mesage manager, registry operation,
resource manager and the cooperative learning model. The common interface provides a standard
framework for communication and messaging across robots. The message manager works like a
master control node for handling the message exchange between system and robots. The registry
manages information about the robots that are availbale on the distributed system. The resource
manager is responsible for handling requests for information from robots. Finally, the cooperative
learning model supports learning and searching for knowledge of grasping strategies for objects. The
platform works with a robot registry database and a repository for grasping point information.
My methodology can be divided into four main parts: object recognition and grasping point detection,
development of the environment for cloud robotics, cooperative learning via cloud robotics, and ex-
perimental evaluation. I will evaluate the system using the Robotics Automation Virtual Environment
provided by OpenRAVE (2013) software for quantitative data. I will also perform a limited qualita-
tive real-world evaluation using a Turtlebot 2 and lightweight manipulator, as shown in Figure 3.2.
This evaluation will focus on elder care assistance in fetching small objects.