ROS Resources: Documentation | Support | Discussion Forum | Service Status | Q&A answers.ros.org

Funded PhD Scholarship – Robots that Learn by Imitating Humans

Applications are invited for a PhD studentship in the Faculty of Computing and Engineering, Ulster University. Successful candidates will enrol in October 2016, or as soon as possible thereafter, on a full- time research programme of up to three years’ duration subject to satisfactory progress, leading to the award of the degree of Doctor of Philosophy.

Brief Description

Continuous advancements in robotics show that robots are being increasingly equipped with complex skills to solve a variety of problems. These skills are the result of research being conducted in laboratories and companies, crafted by developers and tested in sample environments. However, unlike a computer program, a robot has to operate in a world where the possibilities are potentially infinite and where it has to continuously adapt its basic programmed skills to face previously unforeseen situations. Unfortunately there has been no successful development of robots that are autonomously capable of improving or adapting the basic skills with which they were initially equipped – something that comes natural to humans.

In this project we propose to address the above problems by developing a new skills building framework that allows a robot to successfully complete complex tasks by using previously learned primitive actions obtained from a skills library that resides in the Cloud. This library will be populated with primitive actions, such as grasping or object manipulation, from many robot sources, therefore vastly increasing the knowledge available for robots when faced with an unknown complex task. To acquire these primitive actions, an approach known as Dynamic Motion Primitives (DMPs) will be used to imitate the bahaviour of a humans action [1]. This one-shot learning approach will enable robot skills to be derived from observations of a human’s solution to a task, omitting the requirement to analytically decompose and manually program a desired behaviour. The developed framework will combine such primitive actions in a hierarchical manner to accomplish tasks that require a more complex solution.

Consider the following scenario: A robot is requested to put away groceries in a kitchen. The robot may have been previously taught via human demonstration to complete a number of primitive tasks such as how to successfully grasp and move each of the different items, however it may have no prior knowledge of opening the door of the kitchen cupboard where the objects are to be put away. Here the robot would query the proposed skills library in the cloud in order to obtain instruction how to open the door, and then generate a sequence of primitive tasks to execute in order to successfully complete the overall task.

On achieving this, we will further extend the approach to enable multiple robots to co-operate on a single task, utilising cognitive approaches for task allocation based on existing robot skills or a robot’s capability to perform a new skill. We will incorporate recent advances developed in the European project RoboHow [2], where previously learned primitive actions are obtained from a skills library residing in the Cloud, populated with primitive actions from many robot sources.

This project would make use of mobile robots that are available in the ISRC robotics lab [3], in particular the state of the art PR2 mobile manipulator robot [4] and the Shunck manipulator arm robots available in our custom-built manipulator lab. The skills library will use a cloud architecture service such as Microsoft Azure as a platform for robot knowledge sharing.

References

  1. Ijspeert, Auke Jan, Jun Nakanishi, and Stefan Schaal. “Movement imitation with nonlinear dynamical systems in humanoid robots.” Proceedings, IEEE International Conference on Robotics and Automation, Vol. 2, 2002.
  2. http://www.robohow.eu
  3. http://isrc.ulster.ac.uk/Cognitive-Robotics-Team/Home.html
  4. http://www.willowgarage.com/pages/pr2/overview

The studentship will comprise tuition fees and a maintenance award for UK and international candidates. All applicants should hold a first or upper second class honours degree, or its equivalent, in Computer Science or a related discipline. For further information on the application process please visit our website: https://www.ulster.ac.uk/research-and-innovation/phdresearch-degrees/how-to-apply

All applications should be online with all supplementary documents uploaded onto the system. The closing date for applications is 19th August 2016

Interviews will form part of the selection process and are likely to be held during September 2016.

If you have any queries regarding this project, do not hesitate to contact Dr. Bryan Gardiner b.gardiner@ulster.ac.uk