A social networking-enabled framework for autonomous robot skill development

Publication Type:
Thesis
Issue Date:
2016
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf274.25 kB
Adobe PDF
02whole.pdf2.7 MB
Adobe PDF
Intelligent service robots will need to adapt to unforeseen situations when performing tasks for humans. To do this, they will be expected to continuously develop new skills. Existing frameworks that address robot learning of new skills for a particular task often follow a task-specific design approach. Task-specific design is unable to support robots to adapt new skills to new tasks. This is largely due to the inability of skill specification in task-specific design to be extended or to be easily changed. This dissertation provides an innovative task-independent framework that allows robots to develop new skills on their own. The idea is to create an online social network platform called Numbots that enables robots to learn new skills autonomously from their social circles. This platform integrates a state-of-the-art approach to learning from experience, called Constructing Skill Trees (CST), with a state-of-the-art framework for knowledge sharing, called RoboEarth. Based on this integration, a new logic model for online Robot-Robot Interaction (RRI) is developed. The principal focus of this dissertation is the analysis of, and solutions to three underlying technical challenges required to achieve the RRI model: (i) skill representation; (ii) autonomous skill recognition and sharing; and (iii) skill transfer. We focus on motion skills required to interact with and manipulate objects where a robot performs a series of motions to attain a goal given by humans. Skills formalise robot activities, which may involve an object (for example, kicking a ball, lifting a box, or passing a bottle of water to a person). Skills may also include robot activities that do not involve objects (for example, raising hands or walking forward). The first challenge concerns how to create a new skill representation that can represent robot skills independently of robot species, tasks and environments. We develop a generic robot skill representation, which characterises three key dimensions of a robot skill in the focused domain: the changing relationship, the spatial relationship and the temporal relationship between the robot and a possible object. The new representation takes a spatial-temporal perspective similar to that found in RoboEarth, and uses the concepts of “agent space” and “object space” from the CST approach. The second challenge concerns how to enable robots to autonomously recognise and share their experiences with other robots that are in their social network. We propose an effect-based skill recognition mechanism that enables robots to recognise skills based on the effects that result from their action. We introduce two types of autonomous skill recognition: (i) recognition of a chain of existing skill primitives; (ii) recognition of a chain of unknown skills. All recognised skills are generalised and packed into a JSON file to share across Numbots. The third challenge is how to enable shared generic robot skills to be interpreted by a robot learner for its own problem solving. We introduce an effect-based skill transfer mechanism, an algorithm to decompose and customise the downloaded generic robot skill into a set of executable action commands for the robot learner's own problem solving. After the introduction of three technical challenges of the RRI model and our solutions, a simulation is undertaken. It demonstrates that a skill recognised and shared by a PR2 robot can be reused and transferred by a NAO robot for a different problem solving. In addition, we also provide a series of comparisons with RoboEarth with a use case study “ServeADrink” to demonstrate the key advantages of the newly created generic robot skill representation over the limited skill representation in RoboEarth. Even though implementation of Numbots and the RRI model on a real robot remains as future work, the proposed analysis and solutions in this dissertation have demonstrated the potential to enable robots to develop new skills on their own, in the absence of human/robot demonstrators and to perform a task for which the robot was not explicitly programmed.
Please use this identifier to cite or link to this item: