A syntactic approach to robot learning of human tasks from demonstrations
Author(s)
Lee, KyuHwa
Type
Thesis
Abstract
The successful development of general-purpose humanoid robots, in contrast to traditional pre-programmed problem solving machines, has opened a new research area of how a robot could be programmed by an end-user, not engineers, to suit individual needs.
In this respect, Robot Learning from Demonstration has been actively studied, aiming to enable robots learn various tasks from human users. Although much effort has been put, there are many challenges still remaining until the goal is realized. One of the important challenges is the automatic learning of task representations and reuse of the learned tasks, where each task can be expressed as a series of primitive action components. To deal with such challenges, syntactic approaches to task learning and related issues are investigated.
Firstly, efficient goal-oriented task representation methods using stochastic context-free grammars are studied, which enable robots to understand the human's intended actions even in the presence of both observation errors and human execution errors. By exploiting the task knowledge, it is demonstrated that the robot can correctly identify unexpected, out-of-context actions and perform the intended actions under reasonable amount of noise.
Taking a step further, the automatic learning of these task representations from human demonstrations are studied. It is demonstrated throughout the experiments that the robot is able to learn critical task structures and generalize them. This is essential for understanding more complex tasks sharing the same underlying structures.
Following these studies, the unsupervised discovery of the optimal number of primitive action detectors required to represent a task is studied.
Through a diverse set of real-world and simulated experiments that include learning object-related games, postural sequence tasks of dance and surveillance tasks, this thesis demonstrates the effectiveness of syntactic approaches for robot learning from demonstrations.
In this respect, Robot Learning from Demonstration has been actively studied, aiming to enable robots learn various tasks from human users. Although much effort has been put, there are many challenges still remaining until the goal is realized. One of the important challenges is the automatic learning of task representations and reuse of the learned tasks, where each task can be expressed as a series of primitive action components. To deal with such challenges, syntactic approaches to task learning and related issues are investigated.
Firstly, efficient goal-oriented task representation methods using stochastic context-free grammars are studied, which enable robots to understand the human's intended actions even in the presence of both observation errors and human execution errors. By exploiting the task knowledge, it is demonstrated that the robot can correctly identify unexpected, out-of-context actions and perform the intended actions under reasonable amount of noise.
Taking a step further, the automatic learning of these task representations from human demonstrations are studied. It is demonstrated throughout the experiments that the robot is able to learn critical task structures and generalize them. This is essential for understanding more complex tasks sharing the same underlying structures.
Following these studies, the unsupervised discovery of the optimal number of primitive action detectors required to represent a task is studied.
Through a diverse set of real-world and simulated experiments that include learning object-related games, postural sequence tasks of dance and surveillance tasks, this thesis demonstrates the effectiveness of syntactic approaches for robot learning from demonstrations.
Version
Open Access
Date Issued
2013-12
Date Awarded
2014-03
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Demiris, Yiannis
Sponsor
European Union
European Union
Engineering and Physical Sciences Research Council
Grant Number
FP7-ICT-270490
FP7-ICT-248116
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)