Use this URL to cite or link to this record in EThOS:
Title: Towards robot learning of tool manipulation from demonstrations
Author: Wu, Yan
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2013
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Robot learning from demonstrations (RLD), also known as imitation learning, plays an important role in human-robot interactions (HRI) because it creates a userfriendly environment for non-expert end-users to teach a robot new skills. Recent advances in sophisticated humanoid robots provide an opportunity for these robots to manipulate everyday tools. This thesis focuses on some key aspects in enabling humanoid robots to learn manipulation of tools using an RLD approach as a threefold investigation. Firstly, this work investigates the outstanding issues in existing RLD frameworks for applications in a real-world HRI environment and presents an RLD model using template-based approach integrated with an online learning algorithm to address these issues. Further investigations probe into biological systems for inspirations on improving the reusability of learned skills which results in an integrated framework using the proposed RLD model as the base model. Attention is then turned to an important aspect tightly coupled with RLD learning for tool manipulation - recognition of tools. A biologically-inspired framework based on RLD-learned tool affordance is proposed to address the issues of applying to tools the traditional static feature matching approach for object recognition. In a series of experiments, we show that our RLD model is capable of learning skills without demanding the users for repeated demonstrations. The learned skills can be generalised to similar tasks with different environmental constraints and updated as more demonstrations of the skill are presented. We also demonstrate the reusability of the complete framework in comparison to the base model by a realworld HRI application in playing the tic-tac-toe game. Through the experiments on tool recognition using a set of human demonstrations, the benchmarked performance indicates that our proposed tool representation framework is a suitable supplement to the current object recognition models on tools.
Supervisor: Demiris, Yiannis Sponsor: Agency for Science, Technology and Research, Singapore
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available