|
| Speaker: | |
| Department of Artificial Intelligence, School of Computer Science & Engineering, The University of New South Wales | |
| Date: | Tuesday 21st July 1998 |
| Time: | 11:30am |
| Place: | Seminar Room 357, Building E6A, Macquarie University. |
Abstract:
In this presentation, I aim to draw on my experiences in both natural language understanding and robot development.
First, I'll examine four robot experiments, extracting lessons from each. From behaviour-based robots such as my own Madmax, we see the importance of reacting to the real world in real time using independent behaviour modules operating in parallel. From Ian Horswill's Polly, we learn the primacy of good sensory data in moving about in an office environment. David Miller's Scarecrow reminds us that intelligence has a physical aspect, the importance of which has still not been fully realised. Next, two natural language experiments will be examined, again with the aim of finding lessons relevant to the understanding of language for service-oriented planning and motion in the world of offices. From SAVVY, we see that according less value to syntactical structure of utterances and more to pragmatic expectations of the agent leads can lead to reliable, if somewhat inflexible, knowledge structures. From Vere & Bickmore's Homer, we see numerous advantages and difficulties of a complete, language-using integrated reasoner/planner in the service of an active agent. One of greatest challenges is to bridge the gap between language processing/planning/reasoning - essentially high-level, symbolic manipulation - and the lower-level sensory/motor messages of a robot operating in the real world. In the final section, I'll suggest ways in which knowledge structures flowing from language can be made to inform a working robot. My BEELINE experiments show how, given a common representation language for perceptual, active motive and conceptual states, the output of a semantic parser can be made to interact with the operation of a goal-seeking, reactive navigating agent to allow instruction-following.Enquiries: sals@mri.mq.edu.au
| Last modified: July, 1998 |