I will be giving a talk on 23rd Nov, as part of the Hamming Seminar series, where I will attempt to lay out my case for a line of research that I am pursuing with my research group. I welcome you to come, participate in the discussions.
Where are yesterday’s robots of tomorrow? The optimist might answer that some such robots have already begun to enter the realm of the possible, in the form of self-driving cars, robots that can fold towels, take an elevator to fetch you a sandwich, etc. The pessimist would rightly counter that although there are media clips showing robots doing these things, most of these systems are frightfully far from being ready for prime time (do we really expect to be able to take home an Asimo along with our iPhones?). The realist would note that although the pessimist’s objections are valid, a variety of systems ranging from automated vacuum cleaners to unmanned submarines are routinely and gainfully deployed around the world today. Moreover, what about the much larger ecosystem of virtual robots roaming in worlds ranging from eBay and Amazon to NASDAQ?
Siding with the pragmatic realist, I argue that autonomous robots that do succeed and become commonplace will perhaps come in the form of collections of heterogeneous modules that are put together in an ad hoc way to do whatever is the need of the hour. Moreover, people are more likely to accept these robots in mixed-initiative team situations, capable of strategic interaction, rather than as black box unknowns. In this milieu, one technical issue stands out as being particularly thorny – how does one marshal diverse modules into long-lived autonomous systems capable of interacting in such non-trivial ways with a continually changing and rich world?
A proper answer to this question requires a careful look at foundational issues in the science of autonomous decision making and an understanding of the space of implementable models for interactive behaviour. Conceptual advances in this area must also be informed by an empirical science of decisions in boundedly rational agents. One way to do this kind of science is by trying to construct complete agents that act in domains that are rich enough to fully exercise the theoretical framework being developed, while being sufficiently well posed to enable incremental experiments. Two such domains are robotic football and autonomous agents in electronic markets. This talk will provide glimpses into why and how AI researchers get robots to play these games, and how this helps us address the big questions.