Annual Report 2002 - Ãrebro universitet
Annual Report 2002 - Ãrebro universitet
Annual Report 2002 - Ãrebro universitet
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
58 AASS – Center for Applied Autonomous Sensor Systems<br />
2.4.2 Research Projects<br />
Project: Hierarchical Learning of Robot Behaviours<br />
Project leader: Tom Duckett<br />
Project staff: Li Jun<br />
Funding: KK and Faculty funding.<br />
Cooperation: not yet<br />
Synopsis:<br />
This project will develop a generic architecture for learning sensor-motor<br />
behaviours in robotic systems. The robot will develop its own internal<br />
representations through a process of incremental on-line learning. One<br />
application of this technology would be programming by demonstration,<br />
where the human operator simply shows the robot what to do (e.g., ”water<br />
my plants”, ”clean my bathroom”, etc.). The behaviours acquired will be<br />
complex, non-linear and non-achievable by a purely reactive system, so<br />
the learning system must incorporate representations that encode the<br />
temporal and spatial context of the task and environment. A particular<br />
challenge will be to carry out all of the learning in real-time. The project<br />
investigates the implementation of a hierarchy of fast, on-line learning<br />
algorithms that are integrated by layered learning. The techniques<br />
investigated include, but are not limited to, embedded machine learning<br />
algorithms such as self-organisation, recurrent artificial neural networks<br />
and reinforcement learning.<br />
Results in <strong>2002</strong>:<br />
- Analysis and design of a neural network architecture for behaviour learning incorporating<br />
unsupervised and supervised learning. An experimental analysis has been conducted for<br />
the application of mobile robot control.<br />
- Paper published in Swedish Workshop on Autonomous Robots, <strong>2002</strong>.<br />
Future developments:<br />
- The next experiments will investigate alternative applications and platforms, e.g.,<br />
feedback control of a mobile manipulator, and alternative sensory inputs, e.g., vision.<br />
- Adaptation of the acquired behaviours through on-line learning<br />
- Licentiate thesis: expected 2003 (Li Jun)<br />
- PhD-thesis: expected 2005 (Li Jun)